Add file

README.md

@@ -10,359 +10,68 @@ pinned: false
 license: mit
 ---
 
+# 🎙️ VALL‑E‑X_JP-Voice-Cloner
 
-# VALL-E X: Multilingual Text-to-Speech Synthesis and Voice Cloning 🔊
-[![Discord](https://img.shields.io/badge/Discord-%235865F2.svg?style=for-the-badge&logo=discord&logoColor=white)](https://discord.gg/qCBRmAnTxg)
-<br>
-English | [中文](README-ZH.md)
-<br>
-An open source implementation of Microsoft's [VALL-E X](https://arxiv.org/pdf/2303.03926) zero-shot TTS model.<br>
-**We release our trained model to the public for research or application usage.**
+Zero-shot 音声クローンができる日本語対応の音声合成アプリです。  
+1〜3秒の音声サンプルと台本テキストを入力するだけで、  
+**話者の特徴を保持した新しいセリフ音声を生成**できます。  
+モデルは Microsoft の [VALL-E X](https://arxiv.org/pdf/2303.03926) を再現・公開した  
+[Plachtaa/VALL-E-X](https://github.com/Plachtaa/VALL-E-X) に基づいています。
 
-![vallex-framework](/images/vallex_framework.jpg "VALL-E X framework")
-
-VALL-E X is an amazing multilingual text-to-speech (TTS) model proposed by Microsoft. While Microsoft initially publish in their research paper, they did not release any code or pretrained models. Recognizing the potential and value of this technology, our team took on the challenge to reproduce the results and train our own model. We are glad to share our trained VALL-E X model with the community, allowing everyone to experience the power next-generation TTS! 🎧
-<br>
-<br>
-More details about the model are presented in [model card](./model-card.md).
-
-## 📖 Quick Index
-* [🚀 Updates](#-updates)
-* [📢 Features](#-features)
-* [💻 Installation](#-installation)
-* [🎧 Demos](#-demos)
-* [🐍 Usage](#-usage-in-python)
-* [❓ FAQ](#-faq)
-* [🧠 TODO](#-todo)
-
-## 🚀 Updates
-**2023.09.10**
-- Added AR decoder batch decoding for more stable generation result.
-
-**2023.08.30**
-- Replaced EnCodec decoder with Vocos decoder, improved audio quality. (Thanks to [@v0xie](https://github.com/v0xie))
-
-**2023.08.23**
-- Added long text generation.
-
-**2023.08.20**
-- Added [Chinese README](README-ZH.md).
-
-**2023.08.14**
-- Pretrained VALL-E X checkpoint is now released. Download it [here](https://drive.google.com/file/d/10gdQWvP-K_e1undkvv0p2b7SU6I4Egyl/view?usp=sharing)
-
-## 💻 Installation
-### Install with pip, Python 3.10, CUDA 11.7 ~ 12.0, PyTorch 2.0+
-```commandline
-git clone https://github.com/Plachtaa/VALL-E-X.git
-cd VALL-E-X
-pip install -r requirements.txt
-```
-
-> Note: If you want to make prompt, you need to install ffmpeg and add its folder to the environment variable PATH.
-
-When you run the program for the first time, it will automatically download the corresponding model. 
-
-If the download fails and reports an error, please follow the steps below to manually download the model.
-
-(Please pay attention to the capitalization of folders)
-
-1. Check whether there is a `checkpoints` folder in the installation directory. 
-If not, manually create a `checkpoints` folder (`./checkpoints/`) in the installation directory.
-
-2. Check whether there is a `vallex-checkpoint.pt` file in the `checkpoints` folder. 
-If not, please manually download the `vallex-checkpoint.pt` file from [here](https://huggingface.co/Plachta/VALL-E-X/resolve/main/vallex-checkpoint.pt) and put it in the `checkpoints` folder.
-
-3. Check whether there is a `whisper` folder in the installation directory. 
-If not, manually create a `whisper` folder (`./whisper/`) in the installation directory.
-
-4. Check whether there is a `medium.pt` file in the `whisper` folder. 
-If not, please manually download the `medium.pt` file from [here](https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt) and put it in the `whisper` folder.
-
-##  🎧 Demos
-Not ready to set up the environment on your local machine just yet? No problem! We've got you covered with our online demos. You can try out VALL-E X directly on Hugging Face or Google Colab, experiencing the model's capabilities hassle-free!
-<br>
-[![Open in Spaces](https://img.shields.io/badge/🤗-Open%20in%20Spaces-blue.svg)](https://huggingface.co/spaces/Plachta/VALL-E-X)
-[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1yyD_sz531QntLKowMHo-XxorsFBCfKul?usp=sharing)
-
-
-## 📢 Features
-
-VALL-E X comes packed with cutting-edge functionalities:
-
-1. **Multilingual TTS**: Speak in three languages - English, Chinese, and Japanese - with natural and expressive speech synthesis.
-
-2. **Zero-shot Voice Cloning**: Enroll a short 3~10 seconds recording of an unseen speaker, and watch VALL-E X create personalized, high-quality speech that sounds just like them!
-
-<details>
-  <summary><h5>see example</h5></summary>
-
-[prompt.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/a7baa51d-a53a-41cc-a03d-6970f25fcca7)
-
-
-[output.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/b895601a-d126-4138-beff-061aabdc7985)
-
-</details>
-
-3. **Speech Emotion Control**: Experience the power of emotions! VALL-E X can synthesize speech with the same emotion as the acoustic prompt provided, adding an extra layer of expressiveness to your audio.
-
-<details>
-  <summary><h5>see example</h5></summary>
-
-https://github.com/Plachtaa/VALL-E-X/assets/112609742/56fa9988-925e-4757-82c5-83ecb0df6266
-
-
-https://github.com/Plachtaa/VALL-E-X/assets/112609742/699c47a3-d502-4801-8364-bd89bcc0b8f1
-
-</details>
-
-4. **Zero-shot Cross-Lingual Speech Synthesis**: Take monolingual speakers on a linguistic journey! VALL-E X can produce personalized speech in another language without compromising on fluency or accent. Below is a Japanese speaker talk in Chinese & English. 🇯🇵 🗣
-
-<details>
-  <summary><h5>see example</h5></summary>
-
-[jp-prompt.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/ea6e2ee4-139a-41b4-837e-0bd04dda6e19)
-
-
-[en-output.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/db8f9782-923f-425e-ba94-e8c1bd48f207)
-
-
-[zh-output.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/15829d79-e448-44d3-8965-fafa7a3f8c28)
-
-</details>
-
-5. **Accent Control**: Get creative with accents! VALL-E X allows you to experiment with different accents, like speaking Chinese with an English accent or vice versa. 🇨🇳 💬
-
-<details>
-  <summary><h5>see example</h5></summary>
-
-[en-prompt.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/f688d7f6-70ef-46ec-b1cc-355c31e78b3b)
-
-
-[zh-accent-output.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/be59c7ca-b45b-44ca-a30d-4d800c950ccc)
-
-
-[en-accent-output.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/8b4f4f9b-f299-4ea4-a548-137437b71738)
-
-</details>
-
-6. **Acoustic Environment Maintenance**: No need for perfectly clean audio prompts! VALL-E X adapts to the acoustic environment of the input, making speech generation feel natural and immersive.
-
-<details>
-  <summary><h5>see example</h5></summary>
-
-[noise-prompt.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/68986d88-abd0-4d1d-96e4-4f893eb9259e)
-
-
-[noise-output.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/96c4c612-4516-4683-8804-501b70938608)
-
-</details>
-
-
-Explore our [demo page](https://plachtaa.github.io/) for a lot more examples!
-
-## 🐍 Usage in Python
-
-<details open>
-  <summary><h3>🪑 Basics</h3></summary>
-
-```python
-from utils.generation import SAMPLE_RATE, generate_audio, preload_models
-from scipy.io.wavfile import write as write_wav
-from IPython.display import Audio
-
-# download and load all models
-preload_models()
-
-# generate audio from text
-text_prompt = """
-Hello, my name is Nose. And uh, and I like hamburger. Hahaha... But I also have other interests such as playing tactic toast.
-"""
-audio_array = generate_audio(text_prompt)
-
-# save audio to disk
-write_wav("vallex_generation.wav", SAMPLE_RATE, audio_array)
-
-# play text in notebook
-Audio(audio_array, rate=SAMPLE_RATE)
-```
-
-[hamburger.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/578d7bbe-cda9-483e-898c-29646edc8f2e)
-
-</details>
-
-<details open>
-  <summary><h3>🌎 Foreign Language</h3></summary>
-<br>
-This VALL-E X implementation also supports Chinese and Japanese. All three languages have equally awesome performance!
-<br>
-
-```python
-
-text_prompt = """
-    チュソクは私のお気に入りの祭りです。 私は数日間休んで、友人や家族との時間を過ごすことができます。
-"""
-audio_array = generate_audio(text_prompt)
-```
-
-[vallex_japanese.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/ee57a688-3e83-4be5-b0fe-019d16eec51c)
-
-*Note: VALL-E X controls accent perfectly even when synthesizing code-switch text. However, you need to manually denote language of respective sentences (since our g2p tool is rule-base)*
-```python
-text_prompt = """
-    [EN]The Thirty Years' War was a devastating conflict that had a profound impact on Europe.[EN]
-    [ZH]这是历史的开始。 如果您想听更多，请继续。[ZH]
-"""
-audio_array = generate_audio(text_prompt, language='mix')
-```
-
-[vallex_codeswitch.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/d8667abf-bd08-499f-a383-a861d852f98a)
-
-</details>
-
-<details open>
-<summary><h3>📼 Voice Presets</h3></summary>
-  
-VALL-E X provides tens of speaker voices which you can directly used for inference! Browse all voices in the [code](/presets)
-
-> VALL-E X tries to match the tone, pitch, emotion and prosody of a given preset. The model also attempts to preserve music, ambient noise, etc.
-
-```python
-text_prompt = """
-I am an innocent boy with a smoky voice. It is a great honor for me to speak at the United Nations today.
-"""
-audio_array = generate_audio(text_prompt, prompt="dingzhen")
-```
-
-[smoky.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/d3f55732-b1cd-420f-87d6-eab60db14dc5)
-
-</details>
-
-<details open>
-<summary><h3>🎙Voice Cloning</h3></summary>
-  
-VALL-E X supports voice cloning! You can make a voice prompt with any person, character or even your own voice, and use it like other voice presets.<br>
-To make a voice prompt, you need to provide a speech of 3~10 seconds long, as well as the transcript of the speech. 
-You can also leave the transcript blank to let the [Whisper](https://github.com/openai/whisper) model to generate the transcript.
-> VALL-E X tries to match the tone, pitch, emotion and prosody of a given prompt. The model also attempts to preserve music, ambient noise, etc.
-
-```python
-from utils.prompt_making import make_prompt
-
-### Use given transcript
-make_prompt(name="paimon", audio_prompt_path="paimon_prompt.wav",
-                transcript="Just, what was that? Paimon thought we were gonna get eaten.")
-
-### Alternatively, use whisper
-make_prompt(name="paimon", audio_prompt_path="paimon_prompt.wav")
-```
-Now let's try out the prompt we've just made!
-```python
-from utils.generation import SAMPLE_RATE, generate_audio, preload_models
-from scipy.io.wavfile import write as write_wav
-
-# download and load all models
-preload_models()
-
-text_prompt = """
-Hey, Traveler, Listen to this, This machine has taken my voice, and now it can talk just like me!
-"""
-audio_array = generate_audio(text_prompt, prompt="paimon")
-
-write_wav("paimon_cloned.wav", SAMPLE_RATE, audio_array)
-
-```
-
-[paimon_prompt.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/e7922859-9d12-4e2a-8651-e156e4280311)
-
-
-[paimon_cloned.webm](https://github.com/Plachtaa/VALL-E-X/assets/112609742/60d3b7e9-5ead-4024-b499-a897ce5f3d5e)
-
-
-</details>
-
-
-<details open>
-<summary><h3>🎢User Interface</h3></summary>
-
-Not comfortable with codes? No problem! We've also created a user-friendly graphical interface for VALL-E X. It allows you to interact with the model effortlessly, making voice cloning and multilingual speech synthesis a breeze.
-<br>
-You can launch the UI by the following command:
-```commandline
-python -X utf8 launch-ui.py
-```
-</details>
-
-## 🛠️ Hardware and Inference Speed
-
-VALL-E X works well on both CPU and GPU (`pytorch 2.0+`, CUDA 11.7 and CUDA 12.0).
-
-A GPU VRAM of 6GB is enough for running VALL-E X without offloading.
-
-## ⚙️ Details
+---
 
-VALL-E X is similar to [Bark](https://github.com/suno-ai/bark), [VALL-E](https://arxiv.org/abs/2301.02111) and [AudioLM](https://arxiv.org/abs/2209.03143), which generates audio in GPT-style by predicting audio tokens quantized by [EnCodec](https://github.com/facebookresearch/encodec).
-<br>
-Comparing to [Bark](https://github.com/suno-ai/bark):
-- ✔ **Light-weighted**: 3️⃣ ✖ smaller,
-- ✔ **Efficient**: 4️⃣ ✖ faster, 
-- ✔ **Better quality on Chinese & Japanese**
-- ✔ **Cross-lingual speech without foreign accent**
-- ✔ **Easy voice-cloning**
-- ❌ **Less languages**
-- ❌ **No special tokens for music / sound effects**
+## 🐾 特徴
 
-### Supported Languages
+- 🇯🇵 **日本語対応**：日本語音声の入力・出力が可能
+- 🎙️ **Zero-shot Cloning**：3秒の音声と文字起こしで話者再現
+- 📜 **テキスト合成**：好きな台本で喋らせられる
+- 🐱 **カジュアルUI**：誰でも使いやすい Gradio Web UI
 
-| Language | Status |
-| --- | :---: |
-| English (en) | ✅ |
-| Japanese (ja) | ✅ |
-| Chinese, simplified (zh) | ✅ |
+---
 
-## ❓ FAQ
+## 🚀 使い方
 
-#### Where is code for training?
-* [lifeiteng's vall-e](https://github.com/lifeiteng/vall-e) has almost everything. There is no plan to release our training code because there is no difference between lifeiteng's implementation.
+1. 左側のパネルから
+   - クローンしたい話者の音声（WAV）
+   - その文字起こし（必須）
+   - 話させたい台本テキスト（任意）
+   を入力
+2. 「🎙️ 音声生成」ボタンをクリック
+3. 右側に生成音声が再生可能な状態で表示されます🎧
 
-#### Where can I download the model checkpoint?
-* We use `wget` to download the model to directory `./checkpoints/` when you run the program for the first time.
-* If the download fails on the first run, please manually download from [this link](https://huggingface.co/Plachta/VALL-E-X/resolve/main/vallex-checkpoint.pt), and put the file under directory `./checkpoints/`.
+---
 
-#### How much VRAM do I need?
-* 6GB GPU VRAM - Almost all NVIDIA GPUs satisfy the requirement.
+## 💻 動作環境
 
-#### Why the model fails to generate long text?
-* Transformer's computation complexity increases quadratically while the sequence length increases. Hence, all training 
-are kept under 22 seconds. Please make sure the total length of audio prompt and generated audio is less than 22 seconds 
-to ensure acceptable performance. 
+- 推論には CPU でも動作可能ですが、GPU があると高速です
+- 利用している主なライブラリ：
+  - `torch`, `torchaudio`, `encodec`, `gradio`, `pyopenjtalk-prebuilt`, `openai-whisper`, など
 
+---
 
-#### MORE TO BE ADDED...
+## 🔗 モ��ルについて
 
-## 🧠 TODO
-- [x] Add Chinese README
-- [x] Long text generation
-- [x] Replace Encodec decoder with Vocos decoder
-- [ ] Fine-tuning for better voice adaptation
-- [ ] `.bat` scripts for non-python users
-- [ ] To be added...
+このアプリは、[Plachtaa/VALL-E-X](https://github.com/Plachtaa/VALL-E-X) によって公開された  
+MITライセンスのコードおよび学習済みモデル（vallex-checkpoint.pt）を利用しています。  
 
-## 🙏 Appreciation
-- [VALL-E X paper](https://arxiv.org/pdf/2303.03926) for the brilliant idea
-- [lifeiteng's vall-e](https://github.com/lifeiteng/vall-e) for related training code
-- [bark](https://github.com/suno-ai/bark) for the amazing pioneering work in neuro-codec TTS model
+モデルの詳細やアーキテクチャは[こちらのモデルカード](https://github.com/Plachtaa/VALL-E-X/blob/main/model-card.md)をご覧ください。
 
-## ⭐️ Show Your Support
+> This app uses the pretrained VALL-E X model by [Plachtaa](https://github.com/Plachtaa/VALL-E-X), released under the MIT License.
 
-If you find VALL-E X interesting and useful, give us a star on GitHub! ⭐️ It encourages us to keep improving the model and adding exciting features.
+---
 
-## 📜 License
+## 📜 ライセンス
 
-VALL-E X is licensed under the [MIT License](./LICENSE).
+本アプリケーションおよび構成コードは **MIT License** に基づいて公開されています。  
+学習済みモデル・データセットの利用は、各リソースの元ライセンスに従ってください。
 
 ---
 
-Have questions or need assistance? Feel free to [open an issue](https://github.com/Plachtaa/VALL-E-X/issues/new) or join our [Discord](https://discord.gg/qCBRmAnTxg)
+## 🧠 クレジット・参考文献
+
+- [VALL-E X 論文](https://arxiv.org/pdf/2303.03926)
+- [Plachtaa/VALL-E-X](https://github.com/Plachtaa/VALL-E-X)
+- [Facebook EnCodec](https://github.com/facebookresearch/encodec)
+- [OpenAI Whisper](https://github.com/openai/whisper)
 
-Happy voice cloning! 🎤
+---

customs/myvoice.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb79aacf47428c4926f657c8d966597798000218cf4b0d5c3e7b9d2bb7ae21de
+size 19514

customs/prompt_1744971044.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c83e6412f2a30c41150272c16134261aae854dc0c7dbaf1605e3b7ea72c8d5a9
+size 32194

customs/prompt_1744971065.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c83e6412f2a30c41150272c16134261aae854dc0c7dbaf1605e3b7ea72c8d5a9
+size 32194

customs/prompt_1744973467.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27120148b86b6ad3cee8c1171e8133142be429dcc319468e60d4ce311e9775f6
+size 32178

data/__init__.py

@@ -0,0 +1,3 @@
+# from .datamodule import *
+# from .tokenizer import *
+from .collation import *

data/collation.py

@@ -0,0 +1,120 @@
+from pathlib import Path
+from typing import List, Tuple
+
+import numpy as np
+import torch
+
+from utils import SymbolTable
+
+
+class TextTokenCollater:
+    """Collate list of text tokens
+
+    Map sentences to integers. Sentences are padded to equal length.
+    Beginning and end-of-sequence symbols can be added.
+
+    Example:
+        >>> token_collater = TextTokenCollater(text_tokens)
+        >>> tokens_batch, tokens_lens = token_collater(text)
+
+    Returns:
+        tokens_batch: IntTensor of shape (B, L)
+            B: batch dimension, number of input sentences
+            L: length of the longest sentence
+        tokens_lens: IntTensor of shape (B,)
+            Length of each sentence after adding <eos> and <bos>
+            but before padding.
+    """
+
+    def __init__(
+        self,
+        text_tokens: List[str],
+        add_eos: bool = True,
+        add_bos: bool = True,
+        pad_symbol: str = "<pad>",
+        bos_symbol: str = "<bos>",
+        eos_symbol: str = "<eos>",
+    ):
+        self.pad_symbol = pad_symbol
+
+        self.add_eos = add_eos
+        self.add_bos = add_bos
+
+        self.bos_symbol = bos_symbol
+        self.eos_symbol = eos_symbol
+
+        unique_tokens = (
+            [pad_symbol]
+            + ([bos_symbol] if add_bos else [])
+            + ([eos_symbol] if add_eos else [])
+            + sorted(text_tokens)
+        )
+
+        self.token2idx = {token: idx for idx, token in enumerate(unique_tokens)}
+        self.idx2token = [token for token in unique_tokens]
+
+    def index(
+        self, tokens_list: List[str]
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        seqs, seq_lens = [], []
+        for tokens in tokens_list:
+            assert (
+                all([True if s in self.token2idx else False for s in tokens])
+                is True
+            )
+            seq = (
+                ([self.bos_symbol] if self.add_bos else [])
+                + list(tokens)
+                + ([self.eos_symbol] if self.add_eos else [])
+            )
+            seqs.append(seq)
+            seq_lens.append(len(seq))
+
+        max_len = max(seq_lens)
+        for k, (seq, seq_len) in enumerate(zip(seqs, seq_lens)):
+            seq.extend([self.pad_symbol] * (max_len - seq_len))
+
+        tokens = torch.from_numpy(
+            np.array(
+                [[self.token2idx[token] for token in seq] for seq in seqs],
+                dtype=np.int64,
+            )
+        )
+        tokens_lens = torch.IntTensor(seq_lens)
+
+        return tokens, tokens_lens
+
+    def __call__(self, texts: List[str]) -> Tuple[torch.Tensor, torch.Tensor]:
+        tokens_seqs = [[p for p in text] for text in texts]
+        max_len = len(max(tokens_seqs, key=len))
+
+        seqs = [
+            ([self.bos_symbol] if self.add_bos else [])
+            + list(seq)
+            + ([self.eos_symbol] if self.add_eos else [])
+            + [self.pad_symbol] * (max_len - len(seq))
+            for seq in tokens_seqs
+        ]
+
+        tokens_batch = torch.from_numpy(
+            np.array(
+                [seq for seq in seqs],
+                dtype=np.int64,
+            )
+        )
+
+        tokens_lens = torch.IntTensor(
+            [
+                len(seq) + int(self.add_eos) + int(self.add_bos)
+                for seq in tokens_seqs
+            ]
+        )
+
+        return tokens_batch, tokens_lens
+
+
+def get_text_token_collater() -> TextTokenCollater:
+    collater = TextTokenCollater(
+        ['0'], add_bos=False, add_eos=False
+    )
+    return collater

data/datamodule.py

@@ -0,0 +1,419 @@
+# Copyright      2023                          (authors: Feiteng Li)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import torch
+# from icefall.utils import str2bool
+# from lhotse import CutSet, load_manifest_lazy
+# from lhotse.dataset import (
+#     CutConcatenate,
+#     DynamicBucketingSampler,
+#     PrecomputedFeatures,
+#     SingleCutSampler,
+#     SpecAugment,
+# )
+# from lhotse.dataset.input_strategies import OnTheFlyFeatures
+# from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from data.collation import get_text_token_collater
+# from data.dataset import SpeechSynthesisDataset
+from data.fbank import get_fbank_extractor
+from data.input_strategies import PromptedPrecomputedFeatures
+
+# PrecomputedFeatures = PrecomputedFeatures
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+def _get_input_strategy(input_strategy, dataset, cuts):
+    if input_strategy == "PromptedPrecomputedFeatures":
+        return PromptedPrecomputedFeatures(dataset, cuts)
+
+    return eval(input_strategy)()
+
+
+class TtsDataModule:
+    """
+    DataModule for VALL-E TTS experiments.
+    It assumes there is always one train and valid dataloader.
+
+    It contains all the common data pipeline modules used in TTS
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation[not used & tested yet],
+    - augmentation[not used & tested yet],
+    - on-the-fly feature extraction[not used & tested yet]
+
+    This class should be derived for specific corpora used in TTS tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="TTS data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/tokenized"),
+            help="Path to directory with train/valid/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=40.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=10,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=0.1,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=False,
+            help="Whether to drop last batch. Used by sampler.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=8,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=False,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures or PromptedPrecomputedFeatures",
+        )
+
+        group.add_argument(
+            "--dataset",
+            type=str,
+            default="ljspeech",
+            help="--input-strategy PromptedPrecomputedFeatures needs dataset name to prepare prompts.",
+        )
+
+        parser.add_argument(
+            "--text-tokens",
+            type=str,
+            default="data/tokenized/unique_text_tokens.k2symbols",
+            help="Path to the unique text tokens file",
+        )
+
+        parser.add_argument(
+            "--sampling-rate",
+            type=int,
+            default=24000,
+            help="""Audio sampling rate.""",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(
+                f"Time warp factor: {self.args.spec_aug_time_warp_factor}"
+            )
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = SpeechSynthesisDataset(
+                get_text_token_collater(self.args.text_tokens),
+                cut_transforms=transforms,
+                feature_input_strategy=OnTheFlyFeatures(get_fbank_extractor()),
+                feature_transforms=input_transforms,
+            )
+        else:
+            train = SpeechSynthesisDataset(
+                get_text_token_collater(self.args.text_tokens),
+                feature_input_strategy=_get_input_strategy(
+                    self.args.input_strategy, self.args.dataset, cuts_train
+                ),
+                cut_transforms=transforms,
+                feature_transforms=input_transforms,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info(
+                "Using SingleCutSampler and sort by duraton(ascending=True)."
+            )
+            cuts_train = cuts_train.to_eager().sort_by_duration(ascending=True)
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = SpeechSynthesisDataset(
+                get_text_token_collater(self.args.text_tokens),
+                feature_input_strategy=OnTheFlyFeatures(get_fbank_extractor()),
+                cut_transforms=[],
+            )
+        else:
+            validate = SpeechSynthesisDataset(
+                get_text_token_collater(self.args.text_tokens),
+                feature_input_strategy=_get_input_strategy(
+                    self.args.input_strategy, self.args.dataset, cuts_valid
+                ),
+                cut_transforms=[],
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=4,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = SpeechSynthesisDataset(
+            get_text_token_collater(self.args.text_tokens),
+            feature_input_strategy=OnTheFlyFeatures(get_fbank_extractor())
+            if self.args.on_the_fly_feats
+            else _get_input_strategy(
+                self.args.input_strategy, self.args.dataset, cuts
+            ),
+            cut_transforms=[],
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.debug("About to create test dataloader")
+        test_dl = DataLoader(
+            test,
+            batch_size=None,
+            sampler=sampler,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest_lazy(
+            self.args.manifest_dir / "cuts_train.jsonl.gz"
+        )
+
+    @lru_cache()
+    def dev_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_dev.jsonl.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> CutSet:
+        logging.info("About to get test cuts")
+        return load_manifest_lazy(self.args.manifest_dir / "cuts_test.jsonl.gz")

data/dataset.py

@@ -0,0 +1,242 @@
+# Copyright      2023                           (authors: Feiteng Li)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+modified from lhoste.dataset.speech_synthesis.py
+"""
+
+import torch
+import math
+import h5py
+from tokenizers import Tokenizer
+from typing import Union, List
+import numpy as np
+from tqdm import tqdm
+
+_pad        = '_'
+_punctuation = ',.!?-~…'
+_letters = 'NQabdefghijklmnopstuvwxyzɑæʃʑçɯɪɔɛɹðəɫɥɸʊɾʒθβŋɦ⁼ʰ`^#*=ˈˌ→↓↑ '
+symbols = [_pad] + list(_punctuation) + list(_letters)
+
+language_dict = {
+    'en': 0,
+    'zh': 1,
+    'ja': 2,
+}
+def seq2phone(tokens: Union[List, np.ndarray]):
+    """
+    Convert tokenized phoneme ID sequence back to phoneme string
+    :param tokens: phoneme tokens
+    :return: recovered phoneme sequence
+    """
+    phones = "".join([symbols[i] for i in tokens])
+    return phones
+
+class DynamicBatchSampler(torch.utils.data.Sampler):
+    def __init__(self, sampler, num_tokens_fn, num_buckets=100, min_size=0, max_size=1000,
+                 max_tokens=None, max_sentences=None, drop_last=False):
+        """
+
+        :param sampler:
+        :param num_tokens_fn: 根据idx返回样本的长度的函数
+        :param num_buckets: 利用桶原理将相似长度的样本放在一个batchsize中，桶的数量
+        :param min_size: 最小长度的样本， 小于这个值的样本会被过滤掉。 依据这个值来创建样桶
+        :param max_size: 最大长度的样本
+        :param max_sentences: batch_size, 但是这里可以通过max_sentences 和 max_tokens 共同控制最终的大小
+        """
+        super(DynamicBatchSampler, self).__init__(sampler)
+        self.sampler = sampler
+        self.num_tokens_fn = num_tokens_fn
+        self.num_buckets = num_buckets
+
+        self.min_size = min_size
+        self.max_size = max_size
+
+        assert max_size <= max_tokens, "max_size should be smaller than max tokens"
+        assert max_tokens is not None or max_sentences is not None, \
+            "max_tokens and max_sentences should not be null at the same time, please specify one parameter at least"
+        self.max_tokens = max_tokens if max_tokens is not None else float('Inf')
+        self.max_sentences = max_sentences if max_sentences is not None else float('Inf')
+        self.drop_last = drop_last
+
+    def set_epoch(self, epoch):
+        self.sampler.set_epoch(epoch)
+    def is_batch_full(self, num_tokens, batch):
+        if len(batch) == 0:
+            return False
+        if len(batch) == self.max_sentences:
+            return True
+        if num_tokens > self.max_tokens:
+            return True
+        return False
+
+    def __iter__(self):
+        buckets = [[] for _ in range(self.num_buckets)]
+        sample_len = [0] * self.num_buckets
+
+        for idx in self.sampler:
+            idx_length = self.num_tokens_fn(idx)
+            if not (self.min_size <= idx_length <= self.max_size):
+                print("sentence at index {} of size {} exceeds max_tokens, the sentence is ignored".format(idx, idx_length))
+                continue
+
+            index_buckets = math.floor((idx_length - self.min_size) / (self.max_size - self.min_size + 1)
+                                       * self.num_buckets)
+            sample_len[index_buckets] = max(sample_len[index_buckets], idx_length)
+
+            num_tokens = (len(buckets[index_buckets]) + 1) * sample_len[index_buckets]
+            if self.is_batch_full(num_tokens, buckets[index_buckets]):
+                # yield this batch
+                yield buckets[index_buckets]
+                buckets[index_buckets] = []
+                sample_len[index_buckets] = 0
+
+            buckets[index_buckets].append(idx)
+
+        # process left-over
+        leftover_batch = []
+        leftover_sample_len = 0
+        leftover = [idx for bucket in buckets for idx in bucket]
+        for idx in leftover:
+            idx_length = self.num_tokens_fn(idx)
+            leftover_sample_len = max(leftover_sample_len, idx_length)
+            num_tokens = (len(leftover_batch) + 1) * leftover_sample_len
+            if self.is_batch_full(num_tokens, leftover_batch):
+                yield leftover_batch
+                leftover_batch = []
+                leftover_sample_len = 0
+            leftover_batch.append(idx)
+
+        if len(leftover_batch) > 0 and not self.drop_last:
+            yield leftover_batch
+
+    def __len__(self):
+        # we do not know the exactly batch size, so do not call len(dataloader)
+        pass
+
+
+class AudioDataset(torch.utils.data.Dataset):
+    def __init__(self, h5_path, ann_path, tokenizer_path):
+        self.h5_path = h5_path
+        with open(ann_path, 'r', encoding='utf-8') as f:
+            lines = f.readlines()
+        ls = [l.split("|") for l in lines]
+        ls_T = list(zip(*ls))
+        del ls_T[-1]
+        self.h5_paths, self.durations, self.langs, self.texts = \
+            list(ls_T[0]), list(ls_T[1]), list(ls_T[2]), list(ls_T[3])
+        self.durations = [float(dur) for dur in self.durations]
+        self.tokenizer = Tokenizer.from_file(tokenizer_path)
+
+        self._archive = None
+
+    def __len__(self):
+        return len(self.h5_paths)
+
+    def get_dur(self, idx):
+        return self.durations[idx]
+
+    @property
+    def archive(self):
+        if self._archive is None:  # lazy loading here!
+            self._archive = h5py.File(self.h5_path, "r")
+        return self._archive
+    def __getitem__(self, idx):
+        archive = self.archive
+        h5_path = self.h5_paths[idx]
+        sub = archive[h5_path]
+        audio_tokens = sub['audio'][()]
+        phone_tokens = sub['text'][()]
+        dur = self.durations[idx]
+        lang = self.langs[idx]
+        text = self.texts[idx]
+        # tokenization should be done within dataloader
+        phones = seq2phone(phone_tokens)
+        phones = phones.replace(" ", "_")
+        if not len(phones):
+            cptpho_tokens = self.tokenizer.encode(text).ids
+        else:
+            cptpho_tokens = self.tokenizer.encode(phones).ids
+        assert len(cptpho_tokens)
+        return {
+            'utt_id': h5_path,
+            'text': text,
+            'audio': None,
+            'audio_lens': None,
+            'audio_features': audio_tokens,
+            'audio_features_lens': len(audio_tokens.T),
+            'text_tokens': np.array(cptpho_tokens),
+            'text_tokens_lens': len(cptpho_tokens),
+            'language': language_dict[lang],
+        }
+
+def collate(batch):
+    utt_id_s = [b['utt_id'] for b in batch]
+    text_s = [b['text'] for b in batch]
+
+    audio_s = [b['audio'] for b in batch]
+    audio_lens_s = [b['audio_lens'] for b in batch]
+
+    audio_features_lens_s = [b['audio_features_lens'] for b in batch]
+    # create an empty tensor with maximum audio feature length
+    audio_features_s = torch.zeros([len(batch), max(audio_features_lens_s), 8], dtype=torch.int64) - 1 # audio pad with -1
+
+    text_tokens_lens_s = [b['text_tokens_lens'] for b in batch]
+    # create an empty tensor with maximum text tokens length
+    text_tokens_s = torch.zeros([len(batch), max(text_tokens_lens_s)], dtype=torch.int64) + 3 # [PAD] token id 3
+
+    language_s = [b['language'] for b in batch]
+
+    for i, b in enumerate(batch):
+        audio_features = b['audio_features']
+        audio_features_lens = b['audio_features_lens']
+        audio_features_s[i, :audio_features_lens, :] = torch.LongTensor(audio_features.T)
+
+        text_tokens = b['text_tokens']
+        text_tokens_lens = b['text_tokens_lens']
+        text_tokens_s[i, :text_tokens_lens] = torch.LongTensor(text_tokens)
+
+    batch = {
+        'utt_id': utt_id_s,
+        'text': text_s,
+        'audio': audio_s,
+        'audio_lens': audio_lens_s,
+        'audio_features': audio_features_s,
+        'audio_features_lens': torch.LongTensor(np.array(audio_features_lens_s)),
+        'text_tokens': text_tokens_s,
+        'text_tokens_lens': torch.LongTensor(np.array(text_tokens_lens_s)),
+        'languages': torch.LongTensor(np.array(language_s)),
+    }
+    return batch
+
+def create_dataloader(data_dir="/root/valle/egs/mix", n_gpus=1, rank=0, num_workers=0, num_buckets=10, max_duration=120):
+    train_dataset = AudioDataset(h5_path=f"{data_dir}/audio_sum.hdf5",
+                                 ann_path=f"{data_dir}/audio_ann_sum.txt",
+                                 tokenizer_path=f"{data_dir}/bpe_69.json")
+    ran_sampler = torch.utils.data.distributed.DistributedSampler(
+            train_dataset,
+            num_replicas=n_gpus,
+            rank=rank,
+            shuffle=True,
+        )
+    dynamic_sampler = DynamicBatchSampler(ran_sampler, train_dataset.get_dur, num_buckets=num_buckets, max_size=20,
+                                          max_tokens=max_duration)
+
+
+    train_loader = torch.utils.data.DataLoader(train_dataset, num_workers=num_workers, collate_fn=collate,
+                                               batch_sampler=dynamic_sampler)
+
+    return train_loader

data/fbank.py

@@ -0,0 +1,212 @@
+# Copyright      2023                          (authors: Feiteng Li)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import asdict, dataclass
+from typing import Any, Dict, Optional, Union
+
+import numpy as np
+import torch
+# from lhotse.features.base import FeatureExtractor
+# from lhotse.utils import EPSILON, Seconds, compute_num_frames
+from librosa.filters import mel as librosa_mel_fn
+
+
+@dataclass
+class BigVGANFbankConfig:
+    # Spectogram-related part
+    # Note that frame_length and frame_shift will be converted to milliseconds before torchaudio/Kaldi sees them
+    frame_length: Seconds = 1024 / 24000.0
+    frame_shift: Seconds = 256 / 24000.0
+    remove_dc_offset: bool = True
+    round_to_power_of_two: bool = True
+
+    # Fbank-related part
+    low_freq: float = 0.0
+    high_freq: float = 12000.0
+    num_mel_bins: int = 100
+    use_energy: bool = False
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+    @staticmethod
+    def from_dict(data: Dict[str, Any]) -> "BigVGANFbankConfig":
+        return BigVGANFbankConfig(**data)
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+# https://github.com/NVIDIA/BigVGAN
+# bigvgan_24khz_100band https://drive.google.com/drive/folders/1EpxX6AsxjCbbk0mmAhE0td6eYiABr8Oz
+class BigVGANFbank(FeatureExtractor):
+    name = "fbank"
+    config_type = BigVGANFbankConfig
+
+    def __init__(self, config: Optional[Any] = None):
+        super(BigVGANFbank, self).__init__(config)
+        sampling_rate = 24000
+        self.mel_basis = torch.from_numpy(
+            librosa_mel_fn(
+                sampling_rate,
+                1024,
+                self.config.num_mel_bins,
+                self.config.low_freq,
+                self.config.high_freq,
+            ).astype(np.float32)
+        )
+        self.hann_window = torch.hann_window(1024)
+
+    def _feature_fn(self, samples, **kwargs):
+        win_length, n_fft = 1024, 1024
+        hop_size = 256
+        if True:
+            sampling_rate = 24000
+            duration = round(samples.shape[-1] / sampling_rate, ndigits=12)
+            expected_num_frames = compute_num_frames(
+                duration=duration,
+                frame_shift=self.frame_shift,
+                sampling_rate=sampling_rate,
+            )
+            pad_size = (
+                (expected_num_frames - 1) * hop_size
+                + win_length
+                - samples.shape[-1]
+            )
+            assert pad_size >= 0
+
+            y = torch.nn.functional.pad(
+                samples,
+                (0, pad_size),
+                mode="constant",
+            )
+        else:
+            y = torch.nn.functional.pad(
+                samples,
+                (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),
+                mode="reflect",
+            )
+
+        y = y.squeeze(1)
+
+        # complex tensor as default, then use view_as_real for future pytorch compatibility
+        spec = torch.stft(
+            y,
+            n_fft,
+            hop_length=hop_size,
+            win_length=win_length,
+            window=self.hann_window,
+            center=False,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+        spec = torch.view_as_real(spec)
+        spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+
+        spec = torch.matmul(self.mel_basis, spec)
+        spec = spectral_normalize_torch(spec)
+
+        return spec.transpose(2, 1).squeeze(0)
+
+    def extract(
+        self, samples: Union[np.ndarray, torch.Tensor], sampling_rate: int
+    ) -> np.ndarray:
+        assert sampling_rate == 24000
+        params = asdict(self.config)
+        params.update({"sample_frequency": sampling_rate, "snip_edges": False})
+        params["frame_shift"] *= 1000.0
+        params["frame_length"] *= 1000.0
+        if not isinstance(samples, torch.Tensor):
+            samples = torch.from_numpy(samples)
+        # Torchaudio Kaldi feature extractors expect the channel dimension to be first.
+        if len(samples.shape) == 1:
+            samples = samples.unsqueeze(0)
+        features = self._feature_fn(samples, **params).to(torch.float32)
+        return features.numpy()
+
+    @property
+    def frame_shift(self) -> Seconds:
+        return self.config.frame_shift
+
+    def feature_dim(self, sampling_rate: int) -> int:
+        return self.config.num_mel_bins
+
+    @staticmethod
+    def mix(
+        features_a: np.ndarray,
+        features_b: np.ndarray,
+        energy_scaling_factor_b: float,
+    ) -> np.ndarray:
+        return np.log(
+            np.maximum(
+                # protection against log(0); max with EPSILON is adequate since these are energies (always >= 0)
+                EPSILON,
+                np.exp(features_a)
+                + energy_scaling_factor_b * np.exp(features_b),
+            )
+        )
+
+    @staticmethod
+    def compute_energy(features: np.ndarray) -> float:
+        return float(np.sum(np.exp(features)))
+
+
+def get_fbank_extractor() -> BigVGANFbank:
+    return BigVGANFbank(BigVGANFbankConfig())
+
+
+if __name__ == "__main__":
+    extractor = BigVGANFbank(BigVGANFbankConfig())
+
+    samples = torch.from_numpy(np.random.random([1000]).astype(np.float32))
+    samples = torch.clip(samples, -1.0, 1.0)
+    fbank = extractor.extract(samples, 24000.0)
+    print(f"fbank {fbank.shape}")
+
+    from scipy.io.wavfile import read
+
+    MAX_WAV_VALUE = 32768.0
+
+    sampling_rate, samples = read(
+        "egs/libritts/prompts/5639_40744_000000_000002.wav"
+    )
+    print(f"samples: [{samples.min()}, {samples.max()}]")
+    fbank = extractor.extract(samples.astype(np.float32) / MAX_WAV_VALUE, 24000)
+    print(f"fbank {fbank.shape}")
+
+    import matplotlib.pyplot as plt
+
+    _ = plt.figure(figsize=(18, 10))
+    plt.imshow(
+        X=fbank.transpose(1, 0),
+        cmap=plt.get_cmap("jet"),
+        aspect="auto",
+        interpolation="nearest",
+    )
+    plt.gca().invert_yaxis()
+    plt.savefig("egs/libritts/prompts/5639_40744_000000_000002.png")
+    plt.close()
+
+    print("fbank test PASS!")

data/input_strategies.py

@@ -0,0 +1,159 @@
+import random
+from collections import defaultdict
+from concurrent.futures import ThreadPoolExecutor
+from typing import Tuple, Type
+
+# from lhotse import CutSet
+# from lhotse.dataset.collation import collate_features
+# from lhotse.dataset.input_strategies import (
+#     ExecutorType,
+#     PrecomputedFeatures,
+#     _get_executor,
+# )
+# from lhotse.utils import fastcopy
+
+
+class PromptedFeatures:
+    def __init__(self, prompts, features):
+        self.prompts = prompts
+        self.features = features
+
+    def to(self, device):
+        return PromptedFeatures(
+            self.prompts.to(device), self.features.to(device)
+        )
+
+    def sum(self):
+        return self.features.sum()
+
+    @property
+    def ndim(self):
+        return self.features.ndim
+
+    @property
+    def data(self):
+        return (self.prompts, self.features)
+
+
+# class PromptedPrecomputedFeatures(PrecomputedFeatures):
+#     """
+#     :class:`InputStrategy` that reads pre-computed features, whose manifests
+#     are attached to cuts, from disk.
+#
+#     It automatically pads the feature matrices with pre or post feature.
+#
+#     .. automethod:: __call__
+#     """
+#
+#     def __init__(
+#         self,
+#         dataset: str,
+#         cuts: CutSet,
+#         num_workers: int = 0,
+#         executor_type: Type[ExecutorType] = ThreadPoolExecutor,
+#     ) -> None:
+#         super(PromptedPrecomputedFeatures, self).__init__(
+#             num_workers, executor_type
+#         )
+#
+#         self.utt2neighbors = defaultdict(lambda: [])
+#
+#         if dataset.lower() == "libritts":
+#             # 909_131041_000013_000002
+#             # 909_131041_000013_000003
+#             speaker2utts = defaultdict(lambda: [])
+#
+#             utt2cut = {}
+#             for cut in cuts:
+#                 speaker = cut.supervisions[0].speaker
+#                 speaker2utts[speaker].append(cut.id)
+#                 utt2cut[cut.id] = cut
+#
+#             for spk in speaker2utts:
+#                 uttids = sorted(speaker2utts[spk])
+#                 # Using the property of sorted keys to find previous utterance
+#                 # The keys has structure speaker_book_x_y e.g. 1089_134691_000004_000001
+#                 if len(uttids) == 1:
+#                     self.utt2neighbors[uttids[0]].append(utt2cut[uttids[0]])
+#                     continue
+#
+#                 utt2prevutt = dict(zip(uttids, [uttids[1]] + uttids[:-1]))
+#                 utt2postutt = dict(zip(uttids[:-1], uttids[1:]))
+#
+#                 for utt in utt2prevutt:
+#                     self.utt2neighbors[utt].append(utt2cut[utt2prevutt[utt]])
+#
+#                 for utt in utt2postutt:
+#                     self.utt2neighbors[utt].append(utt2cut[utt2postutt[utt]])
+#         elif dataset.lower() == "ljspeech":
+#             utt2cut = {}
+#             uttids = []
+#             for cut in cuts:
+#                 uttids.append(cut.id)
+#                 utt2cut[cut.id] = cut
+#
+#             if len(uttids) == 1:
+#                 self.utt2neighbors[uttids[0]].append(utt2cut[uttids[0]])
+#             else:
+#                 # Using the property of sorted keys to find previous utterance
+#                 # The keys has structure: LJ001-0010
+#                 utt2prevutt = dict(zip(uttids, [uttids[1]] + uttids[:-1]))
+#                 utt2postutt = dict(zip(uttids[:-1], uttids[1:]))
+#
+#                 for utt in utt2postutt:
+#                     postutt = utt2postutt[utt]
+#                     if utt[:5] == postutt[:5]:
+#                         self.utt2neighbors[utt].append(utt2cut[postutt])
+#
+#                 for utt in utt2prevutt:
+#                     prevutt = utt2prevutt[utt]
+#                     if utt[:5] == prevutt[:5] or not self.utt2neighbors[utt]:
+#                         self.utt2neighbors[utt].append(utt2cut[prevutt])
+#         else:
+#             raise ValueError
+#
+#     def __call__(
+#         self, cuts: CutSet
+#     ) -> Tuple[PromptedFeatures, PromptedFeatures]:
+#         """
+#         Reads the pre-computed features from disk/other storage.
+#         The returned shape is``(B, T, F) => (batch_size, num_frames, num_features)``.
+#
+#         :return: a tensor with collated features, and a tensor of ``num_frames`` of each cut before padding.
+#         """
+#         features, features_lens = collate_features(
+#             cuts,
+#             executor=_get_executor(
+#                 self.num_workers, executor_type=self._executor_type
+#             ),
+#         )
+#
+#         prompts_cuts = []
+#         for k, cut in enumerate(cuts):
+#             prompts_cut = random.choice(self.utt2neighbors[cut.id])
+#             prompts_cuts.append(fastcopy(prompts_cut, id=f"{cut.id}-{str(k)}"))
+#
+#         mini_duration = min([cut.duration for cut in prompts_cuts] + [3.0])
+#         # prompts_cuts = CutSet.from_cuts(prompts_cuts).truncate(
+#         #     max_duration=mini_duration,
+#         #     offset_type="random",
+#         #     preserve_id=True,
+#         # )
+#         prompts_cuts = CutSet(
+#             cuts={k: cut for k, cut in enumerate(prompts_cuts)}
+#         ).truncate(
+#             max_duration=mini_duration,
+#             offset_type="random",
+#             preserve_id=False,
+#         )
+#
+#         prompts, prompts_lens = collate_features(
+#             prompts_cuts,
+#             executor=_get_executor(
+#                 self.num_workers, executor_type=self._executor_type
+#             ),
+#         )
+#
+#         return PromptedFeatures(prompts, features), PromptedFeatures(
+#             prompts_lens, features_lens
+#         )

data/tokenizer.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+# Copyright    2023                            (authors: Feiteng Li)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+from dataclasses import asdict, dataclass
+from typing import Any, Dict, List, Optional, Pattern, Union
+
+import numpy as np
+import torch
+import torchaudio
+from encodec import EncodecModel
+from encodec.utils import convert_audio
+
+try:
+    from pypinyin import Style, pinyin
+    from pypinyin.style._utils import get_finals, get_initials
+except Exception:
+    pass
+
+
+def remove_encodec_weight_norm(model):
+    from encodec.modules import SConv1d
+    from encodec.modules.seanet import SConvTranspose1d, SEANetResnetBlock
+    from torch.nn.utils import remove_weight_norm
+
+    encoder = model.encoder.model
+    for key in encoder._modules:
+        if isinstance(encoder._modules[key], SEANetResnetBlock):
+            remove_weight_norm(encoder._modules[key].shortcut.conv.conv)
+            block_modules = encoder._modules[key].block._modules
+            for skey in block_modules:
+                if isinstance(block_modules[skey], SConv1d):
+                    remove_weight_norm(block_modules[skey].conv.conv)
+        elif isinstance(encoder._modules[key], SConv1d):
+            remove_weight_norm(encoder._modules[key].conv.conv)
+
+    decoder = model.decoder.model
+    for key in decoder._modules:
+        if isinstance(decoder._modules[key], SEANetResnetBlock):
+            remove_weight_norm(decoder._modules[key].shortcut.conv.conv)
+            block_modules = decoder._modules[key].block._modules
+            for skey in block_modules:
+                if isinstance(block_modules[skey], SConv1d):
+                    remove_weight_norm(block_modules[skey].conv.conv)
+        elif isinstance(decoder._modules[key], SConvTranspose1d):
+            remove_weight_norm(decoder._modules[key].convtr.convtr)
+        elif isinstance(decoder._modules[key], SConv1d):
+            remove_weight_norm(decoder._modules[key].conv.conv)
+
+
+class AudioTokenizer:
+    """EnCodec audio."""
+
+    def __init__(
+        self,
+        device: Any = None,
+    ) -> None:
+        # Instantiate a pretrained EnCodec model
+        model = EncodecModel.encodec_model_24khz()
+        model.set_target_bandwidth(6.0)
+        remove_encodec_weight_norm(model)
+
+        if not device:
+            device = torch.device("cpu")
+            if torch.cuda.is_available():
+                device = torch.device("cuda:0")
+            if torch.backends.mps.is_available():
+                device = torch.device("mps")
+
+        self._device = device
+
+        self.codec = model.to(device)
+        self.sample_rate = model.sample_rate
+        self.channels = model.channels
+
+    @property
+    def device(self):
+        return self._device
+
+    def encode(self, wav: torch.Tensor) -> torch.Tensor:
+        return self.codec.encode(wav.to(self.device))
+
+    def decode(self, frames: torch.Tensor) -> torch.Tensor:
+        return self.codec.decode(frames)
+
+
+def tokenize_audio(tokenizer: AudioTokenizer, audio):
+    # Load and pre-process the audio waveform
+    if isinstance(audio, str):
+        wav, sr = torchaudio.load(audio)
+    else:
+        wav, sr = audio
+    wav = convert_audio(wav, sr, tokenizer.sample_rate, tokenizer.channels)
+    wav = wav.unsqueeze(0)
+
+    # Extract discrete codes from EnCodec
+    with torch.no_grad():
+        encoded_frames = tokenizer.encode(wav)
+    return encoded_frames
+
+
+if __name__ == "__main__":
+    model = EncodecModel.encodec_model_24khz()
+    model.set_target_bandwidth(6.0)
+
+    samples = torch.from_numpy(np.random.random([4, 1, 1600])).type(
+        torch.float32
+    )
+    codes_raw = model.encode(samples)
+
+    remove_encodec_weight_norm(model)
+    codes_norm = model.encode(samples)
+
+    assert torch.allclose(codes_raw[0][0], codes_norm[0][0])

models/__init__.py

@@ -0,0 +1,136 @@
+import argparse
+
+import torch.nn as nn
+# from icefall.utils import AttributeDict, str2bool
+
+from .macros import (
+    NUM_AUDIO_TOKENS,
+    NUM_MEL_BINS,
+    NUM_SPEAKER_CLASSES,
+    NUM_TEXT_TOKENS,
+    SPEAKER_EMBEDDING_DIM,
+)
+from .transformer import Transformer
+from .vallex import VALLE, VALLF
+from .visualizer import visualize
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--model-name",
+        type=str,
+        default="VALL-E",
+        help="VALL-E, VALL-F, Transformer.",
+    )
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=1024,
+        help="Embedding dimension in the decoder model.",
+    )
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=16,
+        help="Number of attention heads in the Decoder layers.",
+    )
+    parser.add_argument(
+        "--num-decoder-layers",
+        type=int,
+        default=12,
+        help="Number of Decoder layers.",
+    )
+    parser.add_argument(
+        "--scale-factor",
+        type=float,
+        default=1.0,
+        help="Model scale factor which will be assigned different meanings in different models.",
+    )
+    parser.add_argument(
+        "--norm-first",
+        type=bool,
+        default=True,
+        help="Pre or Post Normalization.",
+    )
+    parser.add_argument(
+        "--add-prenet",
+        type=bool,
+        default=False,
+        help="Whether add PreNet after Inputs.",
+    )
+
+    # VALL-E & F
+    parser.add_argument(
+        "--prefix-mode",
+        type=int,
+        default=1,
+        help="The mode for how to prefix VALL-E NAR Decoder, "
+        "0: no prefix, 1: 0 to random, 2: random to random, 4: chunk of pre or post utterance.",
+    )
+    parser.add_argument(
+        "--share-embedding",
+        type=bool,
+        default=True,
+        help="Share the parameters of the output projection layer with the parameters of the acoustic embedding.",
+    )
+    parser.add_argument(
+        "--prepend-bos",
+        type=bool,
+        default=False,
+        help="Whether prepend <BOS> to the acoustic tokens -> AR Decoder inputs.",
+    )
+    parser.add_argument(
+        "--num-quantizers",
+        type=int,
+        default=8,
+        help="Number of Audio/Semantic quantization layers.",
+    )
+
+    # Transformer
+    parser.add_argument(
+        "--scaling-xformers",
+        type=bool,
+        default=False,
+        help="Apply Reworked Conformer scaling on Transformers.",
+    )
+
+
+def get_model(params) -> nn.Module:
+    if params.model_name.lower() in ["vall-f", "vallf"]:
+        model = VALLF(
+            params.decoder_dim,
+            params.nhead,
+            params.num_decoder_layers,
+            norm_first=params.norm_first,
+            add_prenet=params.add_prenet,
+            prefix_mode=params.prefix_mode,
+            share_embedding=params.share_embedding,
+            nar_scale_factor=params.scale_factor,
+            prepend_bos=params.prepend_bos,
+            num_quantizers=params.num_quantizers,
+        )
+    elif params.model_name.lower() in ["vall-e", "valle"]:
+        model = VALLE(
+            params.decoder_dim,
+            params.nhead,
+            params.num_decoder_layers,
+            norm_first=params.norm_first,
+            add_prenet=params.add_prenet,
+            prefix_mode=params.prefix_mode,
+            share_embedding=params.share_embedding,
+            nar_scale_factor=params.scale_factor,
+            prepend_bos=params.prepend_bos,
+            num_quantizers=params.num_quantizers,
+        )
+    else:
+        assert params.model_name in ["Transformer"]
+        model = Transformer(
+            params.decoder_dim,
+            params.nhead,
+            params.num_decoder_layers,
+            norm_first=params.norm_first,
+            add_prenet=params.add_prenet,
+            scaling_xformers=params.scaling_xformers,
+        )
+
+    return model

models/macros.py

@@ -0,0 +1,11 @@
+# Text
+NUM_TEXT_TOKENS = 2048
+
+# Audio
+NUM_AUDIO_TOKENS = 1024  # EnCodec RVQ bins
+NUM_MEL_BINS = 100  # BigVGAN bigvgan_24khz_100band
+
+
+# Speaker
+NUM_SPEAKER_CLASSES = 4096
+SPEAKER_EMBEDDING_DIM = 64

models/transformer.py

@@ -0,0 +1,394 @@
+# Copyright    2023                             (authors: Feiteng Li)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partial
+from typing import Any, Dict, List, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from icefall.utils import make_pad_mask
+# from torchmetrics.classification import BinaryAccuracy
+
+from models.vallex import Transpose
+from modules.embedding import SinePositionalEmbedding, TokenEmbedding
+from modules.scaling import BalancedDoubleSwish, ScaledLinear
+from modules.transformer import (
+    BalancedBasicNorm,
+    IdentityNorm,
+    TransformerDecoderLayer,
+    TransformerEncoder,
+    TransformerEncoderLayer,
+)
+
+from .macros import NUM_MEL_BINS, NUM_TEXT_TOKENS
+from .visualizer import visualize
+
+IdentityNorm = IdentityNorm
+
+
+class Transformer(nn.Module):
+    """It implements seq2seq Transformer TTS for debug(No StopPredictor and SpeakerEmbeding)
+    Neural Speech Synthesis with Transformer Network
+    https://arxiv.org/abs/1809.08895
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        num_layers: int,
+        norm_first: bool = True,
+        add_prenet: bool = False,
+        scaling_xformers: bool = False,
+    ):
+        """
+        Args:
+          d_model:
+            The number of expected features in the input (required).
+          nhead:
+            The number of heads in the multiheadattention models (required).
+          num_layers:
+            The number of sub-decoder-layers in the decoder (required).
+        """
+        super().__init__()
+        self.text_embedding = TokenEmbedding(d_model, NUM_TEXT_TOKENS)  # W_x
+
+        if add_prenet:
+            self.encoder_prenet = nn.Sequential(
+                Transpose(),
+                nn.Conv1d(d_model, d_model, kernel_size=5, padding="same"),
+                nn.BatchNorm1d(d_model),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                nn.Conv1d(d_model, d_model, kernel_size=5, padding="same"),
+                nn.BatchNorm1d(d_model),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                nn.Conv1d(d_model, d_model, kernel_size=5, padding="same"),
+                nn.BatchNorm1d(d_model),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                Transpose(),
+                nn.Linear(d_model, d_model),
+            )
+
+            self.decoder_prenet = nn.Sequential(
+                nn.Linear(NUM_MEL_BINS, 256),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                nn.Linear(256, 256),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                nn.Linear(256, d_model),
+            )
+
+            assert scaling_xformers is False  # TODO: update this block
+        else:
+            self.encoder_prenet = nn.Identity()
+            if scaling_xformers:
+                self.decoder_prenet = ScaledLinear(NUM_MEL_BINS, d_model)
+            else:
+                self.decoder_prenet = nn.Linear(NUM_MEL_BINS, d_model)
+
+        self.encoder_position = SinePositionalEmbedding(
+            d_model,
+            dropout=0.1,
+            scale=False,
+        )
+        self.decoder_position = SinePositionalEmbedding(
+            d_model, dropout=0.1, scale=False
+        )
+
+        if scaling_xformers:
+            self.encoder = TransformerEncoder(
+                TransformerEncoderLayer(
+                    d_model,
+                    nhead,
+                    dim_feedforward=d_model * 4,
+                    dropout=0.1,
+                    batch_first=True,
+                    norm_first=norm_first,
+                    linear1_self_attention_cls=ScaledLinear,
+                    linear2_self_attention_cls=partial(
+                        ScaledLinear, initial_scale=0.01
+                    ),
+                    linear1_feedforward_cls=ScaledLinear,
+                    linear2_feedforward_cls=partial(
+                        ScaledLinear, initial_scale=0.01
+                    ),
+                    activation=partial(
+                        BalancedDoubleSwish,
+                        channel_dim=-1,
+                        max_abs=10.0,
+                        min_prob=0.25,
+                    ),
+                    layer_norm_cls=IdentityNorm,
+                ),
+                num_layers=num_layers,
+                norm=BalancedBasicNorm(d_model) if norm_first else None,
+            )
+
+            self.decoder = nn.TransformerDecoder(
+                TransformerDecoderLayer(
+                    d_model,
+                    nhead,
+                    dim_feedforward=d_model * 4,
+                    dropout=0.1,
+                    batch_first=True,
+                    norm_first=norm_first,
+                    linear1_self_attention_cls=ScaledLinear,
+                    linear2_self_attention_cls=partial(
+                        ScaledLinear, initial_scale=0.01
+                    ),
+                    linear1_feedforward_cls=ScaledLinear,
+                    linear2_feedforward_cls=partial(
+                        ScaledLinear, initial_scale=0.01
+                    ),
+                    activation=partial(
+                        BalancedDoubleSwish,
+                        channel_dim=-1,
+                        max_abs=10.0,
+                        min_prob=0.25,
+                    ),
+                    layer_norm_cls=IdentityNorm,
+                ),
+                num_layers=num_layers,
+                norm=BalancedBasicNorm(d_model) if norm_first else None,
+            )
+
+            self.predict_layer = ScaledLinear(d_model, NUM_MEL_BINS)
+            self.stop_layer = nn.Linear(d_model, 1)
+        else:
+            self.encoder = nn.TransformerEncoder(
+                nn.TransformerEncoderLayer(
+                    d_model,
+                    nhead,
+                    dim_feedforward=d_model * 4,
+                    activation=F.relu,
+                    dropout=0.1,
+                    batch_first=True,
+                    norm_first=norm_first,
+                ),
+                num_layers=num_layers,
+                norm=nn.LayerNorm(d_model) if norm_first else None,
+            )
+
+            self.decoder = nn.TransformerDecoder(
+                nn.TransformerDecoderLayer(
+                    d_model,
+                    nhead,
+                    dim_feedforward=d_model * 4,
+                    activation=F.relu,
+                    dropout=0.1,
+                    batch_first=True,
+                    norm_first=norm_first,
+                ),
+                num_layers=num_layers,
+                norm=nn.LayerNorm(d_model) if norm_first else None,
+            )
+
+            self.predict_layer = nn.Linear(d_model, NUM_MEL_BINS)
+            self.stop_layer = nn.Linear(d_model, 1)
+
+        self.stop_accuracy_metric = BinaryAccuracy(
+            threshold=0.5, multidim_average="global"
+        )
+
+    #     self.apply(self._init_weights)
+
+    # def _init_weights(self, module):
+    #     if isinstance(module, (nn.Linear)):
+    #         module.weight.data.normal_(mean=0.0, std=0.02)
+    #         if isinstance(module, nn.Linear) and module.bias is not None:
+    #             module.bias.data.zero_()
+    #     elif isinstance(module, nn.LayerNorm):
+    #         module.bias.data.zero_()
+    #         module.weight.data.fill_(1.0)
+    #     elif isinstance(module, nn.Embedding):
+    #         module.weight.data.normal_(mean=0.0, std=0.02)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: torch.Tensor,
+        y_lens: torch.Tensor,
+        reduction: str = "sum",
+        train_stage: int = 0,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Union[torch.Tensor, None]]:
+        """
+        Args:
+          x:
+            A 2-D tensor of shape (N, S).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of tokens in `x`
+            before padding.
+          y:
+            A 3-D tensor of shape (N, T, 8).
+          y_lens:
+            A 1-D tensor of shape (N,). It contains the number of tokens in `x`
+            before padding.
+          train_stage:
+            Not used in this model.
+        Returns:
+          Return the predicted audio code matrix, cross-entropy loss and Top-10 accuracy.
+        """
+        del train_stage
+
+        assert x.ndim == 2, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.ndim == 3, y.shape
+        assert y_lens.ndim == 1, y_lens.shape
+
+        assert torch.all(x_lens > 0)
+
+        # NOTE: x has been padded in TextTokenCollater
+        x_mask = make_pad_mask(x_lens).to(x.device)
+
+        x = self.text_embedding(x)
+        x = self.encoder_prenet(x)
+        x = self.encoder_position(x)
+        x = self.encoder(x, src_key_padding_mask=x_mask)
+
+        total_loss, metrics = 0.0, {}
+
+        y_mask = make_pad_mask(y_lens).to(y.device)
+        y_mask_float = y_mask.type(torch.float32)
+        data_mask = 1.0 - y_mask_float.unsqueeze(-1)
+
+        # Training
+        # AR Decoder
+        def pad_y(y):
+            y = F.pad(y, (0, 0, 1, 0, 0, 0), value=0).detach()
+            # inputs, targets
+            return y[:, :-1], y[:, 1:]
+
+        y, targets = pad_y(y * data_mask)  # mask padding as zeros
+
+        y_emb = self.decoder_prenet(y)
+        y_pos = self.decoder_position(y_emb)
+
+        y_len = y_lens.max()
+        tgt_mask = torch.triu(
+            torch.ones(y_len, y_len, device=y.device, dtype=torch.bool),
+            diagonal=1,
+        )
+        y_dec = self.decoder(
+            y_pos,
+            x,
+            tgt_mask=tgt_mask,
+            memory_key_padding_mask=x_mask,
+        )
+
+        predict = self.predict_layer(y_dec)
+        # loss
+        total_loss = F.mse_loss(predict, targets, reduction=reduction)
+
+        logits = self.stop_layer(y_dec).squeeze(-1)
+        stop_loss = F.binary_cross_entropy_with_logits(
+            logits,
+            y_mask_float.detach(),
+            weight=1.0 + y_mask_float.detach() * 4.0,
+            reduction=reduction,
+        )
+        metrics["stop_loss"] = stop_loss.detach()
+
+        stop_accuracy = self.stop_accuracy_metric(
+            (torch.sigmoid(logits) >= 0.5).type(torch.int64),
+            y_mask.type(torch.int64),
+        )
+        # icefall MetricsTracker.norm_items()
+        metrics["stop_accuracy"] = stop_accuracy.item() * y_lens.sum().type(
+            torch.float32
+        )
+
+        return ((x, predict), total_loss + 100.0 * stop_loss, metrics)
+
+    def inference(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: Any = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 2-D tensor of shape (1, S).
+          x_lens:
+            A 1-D tensor of shape (1,). It contains the number of tokens in `x`
+            before padding.
+        Returns:
+          Return the predicted audio code matrix and cross-entropy loss.
+        """
+        assert x.ndim == 2, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+
+        assert torch.all(x_lens > 0)
+
+        x_mask = make_pad_mask(x_lens).to(x.device)
+
+        x = self.text_embedding(x)
+        x = self.encoder_prenet(x)
+        x = self.encoder_position(x)
+        x = self.encoder(x, src_key_padding_mask=x_mask)
+
+        x_mask = make_pad_mask(x_lens).to(x.device)
+
+        # AR Decoder
+        # TODO: Managing decoder steps avoid repetitive computation
+        y = torch.zeros(
+            [x.shape[0], 1, NUM_MEL_BINS], dtype=torch.float32, device=x.device
+        )
+        while True:
+            y_emb = self.decoder_prenet(y)
+            y_pos = self.decoder_position(y_emb)
+
+            tgt_mask = torch.triu(
+                torch.ones(
+                    y.shape[1], y.shape[1], device=y.device, dtype=torch.bool
+                ),
+                diagonal=1,
+            )
+
+            y_dec = self.decoder(
+                y_pos,
+                x,
+                tgt_mask=tgt_mask,
+                memory_mask=None,
+                memory_key_padding_mask=x_mask,
+            )
+            predict = self.predict_layer(y_dec[:, -1:])
+
+            logits = self.stop_layer(y_dec[:, -1:]) > 0  # sigmoid(0.0) = 0.5
+            if y.shape[1] > x_lens.max() * 10 or all(logits.cpu().numpy()):
+                print(
+                    f"TransformerTTS EOS [Text {x_lens[0]} -> Audio {y.shape[1]}]"
+                )
+                break
+
+            y = torch.concat([y, predict], dim=1)
+
+        return y[:, 1:]
+
+    def visualize(
+        self,
+        predicts: Tuple[torch.Tensor],
+        batch: Dict[str, Union[List, torch.Tensor]],
+        output_dir: str,
+        limit: int = 4,
+    ) -> None:
+        visualize(predicts, batch, output_dir, limit=limit)

models/vallex.py

@@ -0,0 +1,853 @@
+# Copyright    2023                             (authors: Feiteng Li)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import random
+from typing import Dict, Iterator, List, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from icefall.utils import make_pad_mask
+# from torchmetrics.classification import MulticlassAccuracy
+
+from data.input_strategies import PromptedFeatures
+from modules.embedding import SinePositionalEmbedding, TokenEmbedding
+from modules.transformer import (
+    AdaptiveLayerNorm,
+    LayerNorm,
+    TransformerDecoderLayer,
+    TransformerEncoder,
+    TransformerEncoderLayer,
+)
+
+from .macros import NUM_AUDIO_TOKENS, NUM_TEXT_TOKENS
+from .visualizer import visualize
+
+
+class Transpose(nn.Identity):
+    """(N, T, D) -> (N, D, T)"""
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        return input.transpose(1, 2)
+
+
+# NOTE: There are two ways to implement the model
+#       1) [VALL-F] standard TransformerDecoder, use x as memory
+#       2) [VALL-E] modified TransformerDecoder like GPT-x(e.g. causal TransformerEncoder),
+#          use x as the prefix of decoder inputs
+class VALLF(nn.Module):
+    """It implements https://arxiv.org/abs/2301.02111
+    "Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers"
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        num_layers: int,
+        norm_first: bool = True,
+        add_prenet: bool = False,
+        decoder_cls: Union[
+            nn.TransformerDecoder, nn.TransformerEncoder
+        ] = nn.TransformerDecoder,
+        decoder_layer_cls: Union[
+            TransformerDecoderLayer, TransformerEncoderLayer
+        ] = TransformerDecoderLayer,
+        prefix_mode: int = 0,
+        share_embedding: bool = True,
+        nar_scale_factor: float = 1.0,
+        prepend_bos: bool = True,
+        num_quantizers: int = 8,
+    ):
+        """
+        Args:
+          d_model:
+            The number of expected features in the input (required).
+          nhead:
+            The number of heads in the multiheadattention models (required).
+          num_layers:
+            The number of sub-decoder-layers in the decoder (required).
+        """
+        super().__init__()
+        nar_d_model = int(d_model * nar_scale_factor)
+
+        self.ar_text_embedding = TokenEmbedding(d_model, NUM_TEXT_TOKENS)  # W_x
+        self.nar_text_embedding = TokenEmbedding(nar_d_model, NUM_TEXT_TOKENS)
+
+        # ID NUM_AUDIO_TOKENS     -> PAD
+        # ID NUM_AUDIO_TOKENS + 1 -> BOS
+        self.ar_audio_prepend_bos = prepend_bos
+        self.ar_audio_embedding = TokenEmbedding(
+            d_model, NUM_AUDIO_TOKENS + 1 + int(prepend_bos)
+        )
+
+        # PreNet
+        if add_prenet:
+            self.ar_text_prenet = nn.Sequential(
+                Transpose(),
+                nn.Conv1d(d_model, d_model, kernel_size=5, padding="same"),
+                nn.BatchNorm1d(d_model),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                nn.Conv1d(d_model, d_model, kernel_size=5, padding="same"),
+                nn.BatchNorm1d(d_model),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                nn.Conv1d(d_model, d_model, kernel_size=5, padding="same"),
+                nn.BatchNorm1d(d_model),
+                nn.ReLU(),
+                nn.Dropout(0.5),
+                Transpose(),
+                nn.Linear(d_model, d_model),
+            )
+
+            self.ar_audio_prenet = nn.Sequential(
+                nn.Linear(d_model, 256),
+                nn.ReLU(),
+                nn.Dropout(0.25),
+                nn.Linear(256, 256),
+                nn.ReLU(),
+                nn.Dropout(0.25),
+                nn.Linear(256, d_model),
+            )
+        else:
+            self.ar_text_prenet = nn.Identity()
+            self.ar_audio_prenet = nn.Identity()
+
+        self.ar_text_position = SinePositionalEmbedding(
+            d_model,
+            dropout=0.1,
+            scale=False,
+            alpha=True,
+        )
+        self.ar_audio_position = SinePositionalEmbedding(
+            d_model,
+            dropout=0.1,
+            scale=False,
+            alpha=True,
+        )
+
+        self.ar_decoder = decoder_cls(
+            decoder_layer_cls(
+                d_model,
+                nhead,
+                dim_feedforward=d_model * 4,
+                dropout=0.1,
+                batch_first=True,
+                norm_first=norm_first,
+            ),
+            num_layers=num_layers,
+            norm=LayerNorm(d_model) if norm_first else None,
+        )
+        self.ar_predict_layer = nn.Linear(
+            d_model, NUM_AUDIO_TOKENS + 1, bias=False
+        )
+
+        self.rng = random.Random(0)
+        self.num_heads = nhead
+        self.prefix_mode = prefix_mode
+        self.num_quantizers = num_quantizers
+
+        assert num_quantizers >= 1
+        if num_quantizers > 1:
+            self.nar_audio_embeddings = nn.ModuleList(
+                [TokenEmbedding(nar_d_model, NUM_AUDIO_TOKENS + 1)]
+                + [
+                    TokenEmbedding(nar_d_model, NUM_AUDIO_TOKENS)
+                    for i in range(num_quantizers - 1)
+                ]
+            )  # W_a
+
+            # PreNet
+            if add_prenet:
+                self.nar_text_prenet = nn.Sequential(
+                    Transpose(),
+                    nn.Conv1d(
+                        nar_d_model, nar_d_model, kernel_size=5, padding="same"
+                    ),
+                    nn.BatchNorm1d(nar_d_model),
+                    nn.ReLU(),
+                    nn.Dropout(0.5),
+                    nn.Conv1d(
+                        nar_d_model, nar_d_model, kernel_size=5, padding="same"
+                    ),
+                    nn.BatchNorm1d(nar_d_model),
+                    nn.ReLU(),
+                    nn.Dropout(0.5),
+                    nn.Conv1d(
+                        nar_d_model, nar_d_model, kernel_size=5, padding="same"
+                    ),
+                    nn.BatchNorm1d(nar_d_model),
+                    nn.ReLU(),
+                    nn.Dropout(0.5),
+                    Transpose(),
+                    nn.Linear(nar_d_model, nar_d_model),
+                )
+                self.nar_audio_prenet = nn.Sequential(
+                    nn.Linear(nar_d_model, 256),
+                    nn.ReLU(),
+                    nn.Dropout(0.25),
+                    nn.Linear(256, 256),
+                    nn.ReLU(),
+                    nn.Dropout(0.25),
+                    nn.Linear(256, nar_d_model),
+                )
+            else:
+                self.nar_text_prenet = nn.Identity()
+                self.nar_audio_prenet = nn.Identity()
+
+            self.nar_text_position = SinePositionalEmbedding(
+                nar_d_model,
+                dropout=0.0,
+                scale=False,
+                alpha=False,
+            )
+            self.nar_audio_position = SinePositionalEmbedding(
+                nar_d_model,
+                dropout=0.1,
+                scale=False,
+                alpha=False,
+            )
+
+            self.nar_decoder = decoder_cls(
+                decoder_layer_cls(
+                    nar_d_model,
+                    int(nhead * nar_scale_factor),
+                    dim_feedforward=nar_d_model * 4,
+                    dropout=0.1,
+                    batch_first=True,
+                    norm_first=norm_first,
+                    adaptive_layer_norm=True,
+                ),
+                num_layers=int(num_layers * nar_scale_factor),
+                norm=AdaptiveLayerNorm(
+                    nar_d_model, norm=nn.LayerNorm(nar_d_model)
+                )
+                if norm_first
+                else None,
+            )
+            self.nar_predict_layers = nn.ModuleList(
+                [
+                    nn.Linear(nar_d_model, NUM_AUDIO_TOKENS, bias=False)
+                    for i in range(num_quantizers - 1)
+                ]
+            )
+            self.nar_stage_embeddings = nn.ModuleList(
+                [
+                    TokenEmbedding(nar_d_model, 1)
+                    for i in range(num_quantizers - 1)
+                ]
+            )
+
+            if share_embedding:
+                # We share the parameters of the output projection layer with the parameters of the acoustic embedding Wa
+                # NOTE(Feiteng): In the experiment, this undermines accuracy
+                # self.ar_predict_layer.weight = self.ar_audio_embedding.weight
+
+                # We also share the parameters of the acoustic embedding layer and the output prediction layer,
+                # which means the weights of the j-th prediction layer are the same as the (j + 1)-th acoustic embedding layer.
+                for j in range(0, num_quantizers - 2):
+                    self.nar_predict_layers[
+                        j
+                    ].weight = self.nar_audio_embeddings[j + 2].weight
+
+    def stage_parameters(self, stage: int = 1) -> Iterator[nn.Parameter]:
+        assert stage > 0
+        if stage == 1:
+            for name, param in self.named_parameters():
+                if name.startswith("ar_"):
+                    print(f" AR parameter: {name}")
+                    yield param
+
+        if stage == 2:
+            for name, param in self.named_parameters():
+                if name.startswith("nar_"):
+                    print(f"NAR parameter: {name}")
+                    yield param
+
+    def stage_named_parameters(
+        self, stage: int = 1
+    ) -> Iterator[Tuple[str, nn.Parameter]]:
+        assert stage > 0
+        if stage == 1:
+            for pair in self.named_parameters():
+                if pair[0].startswith("ar_"):
+                    yield pair
+
+        if stage == 2:
+            for pair in self.named_parameters():
+                if pair[0].startswith("nar_"):
+                    yield pair
+
+    def pad_y_eos(self, y, y_mask_int, eos_id):
+        targets = F.pad(y, (0, 1), value=0) + eos_id * F.pad(
+            y_mask_int, (0, 1), value=1
+        )
+        # inputs, targets
+        if self.ar_audio_prepend_bos:
+            return (
+                F.pad(targets[:, :-1], (1, 0), value=NUM_AUDIO_TOKENS + 1),
+                targets,
+            )
+
+        return targets[:, :-1], targets[:, 1:]
+
+    def _prepare_prompts(self, y, y_lens, codes, nar_stage, y_prompts_codes, prefix_mode):
+        # 5.1 For the NAR acoustic prompt tokens, we select a random segment waveform of 3 seconds
+        # from the same utterance.
+        # We implement this differently.
+        if prefix_mode == 0:
+            # no prefix
+            prefix_len = 0
+            y_emb = self.nar_audio_embeddings[0](y)
+            for j in range(1, nar_stage):
+                # Formula (4) (5)
+                y_emb = y_emb + self.nar_audio_embeddings[j](codes[..., j])
+        elif prefix_mode == 1:
+            # prefix at begining
+            int_low = (0.25 * y_lens.min()).type(torch.int64).item()
+            prefix_len = torch.randint(0, int_low * 2, size=()).item()
+            prefix_len = min(prefix_len, 225)  # 24000/320 * 3s = 225 frames
+
+            y_prompts = self.nar_audio_embeddings[0](y[:, :prefix_len])
+            y_emb = self.nar_audio_embeddings[0](y[:, prefix_len:])
+            for j in range(1, self.num_quantizers):
+                y_prompts += self.nar_audio_embeddings[j](
+                    codes[:, :prefix_len, j]
+                )
+                if j < nar_stage:
+                    y_emb += self.nar_audio_embeddings[j](
+                        codes[:, prefix_len:, j]
+                    )
+            y_emb = torch.concat([y_prompts, y_emb], axis=1)
+        elif prefix_mode in [2, 4]:
+            if prefix_mode == 2:
+                # random prefix
+                prefix_len = min(225, int(0.25 * y_lens.min().item()))
+
+                y_prompts_codes = []
+                for b in range(codes.shape[0]):
+                    start = self.rng.randint(0, y_lens[b].item() - prefix_len)
+                    y_prompts_codes.append(
+                        torch.clone(codes[b, start : start + prefix_len])
+                    )
+                    codes[
+                        b, start : start + prefix_len, nar_stage
+                    ] = NUM_AUDIO_TOKENS
+                y_prompts_codes = torch.stack(y_prompts_codes, dim=0)
+            else:
+                prefix_len = y_prompts_codes.shape[1]
+
+            y_prompts = self.nar_audio_embeddings[0](y_prompts_codes[..., 0])
+            y_emb = self.nar_audio_embeddings[0](y)
+            for j in range(1, self.num_quantizers):
+                y_prompts += self.nar_audio_embeddings[j](
+                    y_prompts_codes[..., j]
+                )
+                if j < nar_stage:
+                    y_emb += self.nar_audio_embeddings[j](codes[..., j])
+            y_emb = torch.concat([y_prompts, y_emb], axis=1)
+        else:
+            raise ValueError
+
+        return y_emb, prefix_len
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: Union[torch.Tensor, PromptedFeatures],
+        y_lens: Union[torch.Tensor, PromptedFeatures],
+        reduction: str = "sum",
+        train_stage: int = 0,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Union[torch.Tensor, None]]:
+        raise NotImplementedError
+
+    def inference(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: torch.Tensor,
+        enroll_x_lens: Union[torch.Tensor, None] = None,
+        top_k: int = -100,
+        temperature: float = 1.0,
+    ) -> torch.Tensor:
+        raise NotImplementedError
+
+    def visualize(
+        self,
+        predicts: Tuple[torch.Tensor],
+        batch: Dict[str, Union[List, torch.Tensor]],
+        output_dir: str,
+        limit: int = 4,
+    ) -> None:
+        raise NotImplementedError
+
+
+class VALLE(VALLF):
+    """It implements https://arxiv.org/abs/2301.02111
+    "Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers"
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        num_layers: int,
+        norm_first: bool = True,
+        add_prenet: bool = False,
+        prefix_mode: int = 0,
+        share_embedding: bool = True,
+        nar_scale_factor: float = 1.0,
+        **kwargs,
+    ):
+        """
+        Args:
+          d_model:
+            The number of expected features in the input (required).
+          nhead:
+            The number of heads in the multiheadattention models (required).
+          num_layers:
+            The number of sub-decoder-layers in the decoder (required).
+        """
+        super(VALLE, self).__init__(
+            d_model,
+            nhead,
+            num_layers,
+            norm_first=norm_first,
+            add_prenet=add_prenet,
+            decoder_cls=TransformerEncoder,
+            decoder_layer_cls=TransformerEncoderLayer,
+            prefix_mode=prefix_mode,
+            share_embedding=share_embedding,
+            nar_scale_factor=nar_scale_factor,
+            **kwargs,
+        )
+        self.language_ID = {
+            'en': 0,
+            'zh': 1,
+            'ja': 2,
+        }
+        self.ar_language_embedding = TokenEmbedding(d_model, len(self.language_ID))
+        self.nar_language_embedding = TokenEmbedding(d_model, len(self.language_ID))
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: Union[torch.Tensor, PromptedFeatures],
+        y_lens: Union[torch.Tensor, PromptedFeatures],
+        reduction: str = "sum",
+        train_stage: int = 0,
+        **kwargs,
+    ):
+        raise NotImplementedError
+    def inference(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: torch.Tensor,
+        enroll_x_lens: torch.Tensor,
+        top_k: int = -100,
+        temperature: float = 1.0,
+        prompt_language: str = None,
+        text_language: str = None,
+        best_of: int = 1,
+        length_penalty: float = 1.0,
+        return_worst: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 2-D tensor of shape (1, S).
+          x_lens:
+            A 1-D tensor of shape (1,). It contains the number of tokens in `x`
+            before padding.
+          y:
+            A 3-D tensor of shape (1, T, 8).
+          top_k: (`optional`) int
+            The number of highest probability tokens to keep for top-k-filtering. Default to -100.
+          temperature: (`optional`) float
+            The value used to module the next token probabilities. Must be strictly positive. Default to 1.0.
+        Returns:
+          Return the predicted audio code matrix.
+        """
+        assert x.ndim == 2, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.ndim == 3, y.shape
+        assert y.shape[0] == 1, y.shape
+
+        assert torch.all(x_lens > 0)
+
+        # NOTE: x has been padded in TextTokenCollater
+        text = x
+        x = self.ar_text_embedding(text)
+        # Add language embedding
+        prompt_language_id = torch.LongTensor(np.array([self.language_ID[prompt_language]])).to(x.device)
+        if isinstance(text_language, str):
+            text_language_id = torch.LongTensor(np.array([self.language_ID[text_language]])).to(x.device)
+        elif isinstance(text_language, List):
+            text_language_id = torch.LongTensor(np.array([self.language_ID[tl] for tl in text_language])).to(x.device)
+        x[:, :enroll_x_lens, :] += self.ar_language_embedding(prompt_language_id)
+        x[:, enroll_x_lens:, :] += self.ar_language_embedding(text_language_id)
+        x = self.ar_text_prenet(x)
+        x = self.ar_text_position(x)
+
+        text_len = x_lens.max()
+        prompts = y
+        prefix_len = y.shape[1]
+
+        # AR Decoder
+        # TODO: Managing decoder steps avoid repetitive computation
+        y = prompts[..., 0]
+        if self.ar_audio_prepend_bos:
+            y = F.pad(y, (1, 0), value=NUM_AUDIO_TOKENS + 1)
+
+        x_len = x_lens.max()
+        x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
+
+        kv_cache = None
+        use_kv_caching = True
+
+        sum_logprobs = torch.zeros(best_of, device=y.device)  # implement batch decoding here
+        x = x.repeat(best_of, 1, 1)
+        y = y.repeat(best_of, 1)
+        while True:
+            y_emb = self.ar_audio_embedding(y)
+            y_emb = self.ar_audio_prenet(y_emb)
+            y_pos = self.ar_audio_position(y_emb)
+            xy_pos = torch.concat([x, y_pos], dim=1)
+
+            y_len = y.shape[1]
+            x_attn_mask_pad = F.pad(
+                x_attn_mask,
+                (0, y_len),
+                value=True,
+            )
+            y_attn_mask = F.pad(
+                torch.triu(
+                    torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1
+                ),
+                (x_len, 0),
+                value=False,
+            )
+            xy_attn_mask = torch.concat(
+                [x_attn_mask_pad, y_attn_mask], dim=0
+            ).to(y.device)
+
+
+            if use_kv_caching and kv_cache is not None:
+                xy_pos = xy_pos[:, [-1]]
+            else:
+                pass
+
+            xy_dec, kv_cache = self.ar_decoder.infer(
+                xy_pos,
+                mask=xy_attn_mask,
+                past_kv=kv_cache,
+                use_cache=use_kv_caching,
+            )
+            # xy_dec, _ = self.ar_decoder(
+            #     (xy_pos, None),
+            #     mask=xy_attn_mask,
+            # )
+
+            logits = self.ar_predict_layer(xy_dec[:, -1])
+            samples, current_logprobs = topk_sampling(
+                logits, top_k=top_k, top_p=1, temperature=temperature
+            )
+            sum_logprobs += current_logprobs * (y[:, -1] != NUM_AUDIO_TOKENS)
+            samples[y[:, -1] == NUM_AUDIO_TOKENS] = NUM_AUDIO_TOKENS
+            completed = (samples[:, -1] == NUM_AUDIO_TOKENS).all()
+            if (
+                completed
+                or (y.shape[1] - prompts.shape[1]) > x_lens.max() * 16
+            ):
+                if prompts.shape[1] == y.shape[1]:
+                    raise SyntaxError(
+                        "well trained model shouldn't reach here."
+                    )
+                lengths = torch.sum(y != NUM_AUDIO_TOKENS, dim=1)
+                avg_logprobs = sum_logprobs / lengths ** length_penalty
+                # choose the best beam according to sum_logprobs
+                best_beam = y[torch.argmax(avg_logprobs), :]
+                worst_beam = y[torch.argmin(avg_logprobs), :]
+                # strip all eos tokens
+                best_beam = best_beam[best_beam != NUM_AUDIO_TOKENS]
+                worst_beam = worst_beam[worst_beam != NUM_AUDIO_TOKENS]
+                if return_worst:
+                    y = worst_beam.unsqueeze(0)
+                else:
+                    y = best_beam.unsqueeze(0)
+                print(f"VALL-E EOS [{prompts.shape[1]} -> {y.shape[1]}]")
+                break
+
+            y = torch.concat([y, samples], dim=1)
+
+        codes = [y[:, prefix_len + int(self.ar_audio_prepend_bos) :]]
+        if self.num_quantizers == 1:
+            return torch.stack(codes, dim=-1)
+
+        # Non-AR Decoders
+        y_emb = self.nar_audio_embeddings[0](
+            y[:, int(self.ar_audio_prepend_bos) :]
+        )
+
+        if self.prefix_mode in [2, 4]:  # Exclude enrolled_phonemes
+            enrolled_len = enroll_x_lens.max().item()
+            # SOS + Synthesis Text + EOS
+            text = torch.concat(
+                [
+                    text[:, :1],
+                    text[:, enrolled_len - 1 :],
+                ],
+                dim=1,
+            )
+            text_len = text_len - (enrolled_len - 2)
+            assert text.shape[0] == 1
+
+        x = self.nar_text_embedding(text)
+        # Add language embedding
+        prompt_language_id = torch.LongTensor(np.array([self.language_ID[prompt_language]])).to(x.device)
+        if isinstance(text_language, str):
+            text_language_id = torch.LongTensor(np.array([self.language_ID[text_language]])).to(x.device)
+        elif isinstance(text_language, List):
+            text_language_id = torch.LongTensor(np.array([self.language_ID[tl] for tl in text_language])).to(x.device)
+        x[:, :enroll_x_lens, :] += self.nar_language_embedding(prompt_language_id)
+        x[:, enroll_x_lens:, :] += self.nar_language_embedding(text_language_id)
+        x = self.nar_text_prenet(x)
+        x = self.nar_text_position(x)
+
+        if self.prefix_mode == 0:
+            for i, (predict_layer, embedding_layer) in enumerate(
+                zip(
+                    self.nar_predict_layers,
+                    self.nar_audio_embeddings[1:],
+                )
+            ):
+                y_pos = self.nar_audio_prenet(y_emb)
+                y_pos = self.nar_audio_position(y_pos)
+                xy_pos = torch.concat([x, y_pos], dim=1)
+
+                xy_dec, _ = self.nar_decoder(
+                    (xy_pos, self.nar_stage_embeddings[i].weight)
+                )
+                logits = predict_layer(xy_dec[:, text_len + prefix_len :])
+
+                samples = torch.argmax(logits, dim=-1)
+                codes.append(samples)
+
+                if i < self.num_quantizers - 2:
+                    y_emb[:, :prefix_len] += embedding_layer(
+                        prompts[..., i + 1]
+                    )
+                    y_emb[:, prefix_len:] += embedding_layer(samples)
+        else:
+            for j in range(1, self.num_quantizers):
+                y_emb[:, :prefix_len] += self.nar_audio_embeddings[j](
+                    prompts[..., j]
+                )
+
+            for i, (predict_layer, embedding_layer) in enumerate(
+                zip(
+                    self.nar_predict_layers,
+                    self.nar_audio_embeddings[1:],
+                )
+            ):
+                y_pos = self.nar_audio_prenet(y_emb)
+                y_pos = self.nar_audio_position(y_pos)
+                xy_pos = torch.concat([x, y_pos], dim=1)
+
+                xy_dec, _ = self.nar_decoder(
+                    (xy_pos, self.nar_stage_embeddings[i].weight)
+                )
+                logits = predict_layer(xy_dec[:, text_len + prefix_len :])
+
+                samples = torch.argmax(logits, dim=-1)
+                codes.append(samples)
+
+                if i < self.num_quantizers - 2:
+                    y_emb[:, prefix_len:] += embedding_layer(samples)
+
+        assert len(codes) == self.num_quantizers
+        return torch.stack(codes, dim=-1)
+
+    def continual(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 2-D tensor of shape (1, S).
+          x_lens:
+            A 1-D tensor of shape (1,). It contains the number of tokens in `x`
+            before padding.
+          y:
+            A 3-D tensor of shape (1, T, 8).
+        Returns:
+          Return the predicted audio code matrix.
+        """
+        assert x.ndim == 2, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.ndim == 3, y.shape
+        assert y.shape[0] == 1, y.shape
+
+        assert torch.all(x_lens > 0)
+        assert self.num_quantizers == 8
+
+        # NOTE: x has been padded in TextTokenCollater
+        text = x
+        x = self.ar_text_embedding(text)
+        x = self.ar_text_prenet(x)
+        x = self.ar_text_position(x)
+
+        text_len = x_lens.max()
+
+        prefix_len = min(int(y.shape[1] * 0.5), 3 * 75)
+
+        # AR Decoder
+        prompts = y[:, :prefix_len]
+
+        codes = [y[:, prefix_len:, 0]]
+        # Non-AR Decoders
+        x = self.nar_text_embedding(text)
+        x = self.nar_text_prenet(x)
+        x = self.nar_text_position(x)
+
+        y_emb = self.nar_audio_embeddings[0](y[..., 0])
+
+        if self.prefix_mode == 0:
+            for i, (predict_layer, embedding_layer) in enumerate(
+                zip(
+                    self.nar_predict_layers,
+                    self.nar_audio_embeddings[1:],
+                )
+            ):
+                y_pos = self.nar_audio_position(y_emb)
+                y_pos = self.nar_audio_prenet(y_pos)
+                xy_pos = torch.concat([x, y_pos], dim=1)
+
+                xy_dec, _ = self.nar_decoder(
+                    (xy_pos, self.nar_stage_embeddings[i].weight)
+                )
+                logits = predict_layer(xy_dec[:, text_len + prefix_len :])
+
+                samples = torch.argmax(logits, dim=-1)
+                codes.append(samples)
+
+                if i < 6:
+                    y_emb[:, :prefix_len] += embedding_layer(
+                        prompts[..., i + 1]
+                    )
+                    y_emb[:, prefix_len:] += embedding_layer(samples)
+        else:
+            for j in range(1, 8):
+                y_emb[:, :prefix_len] += self.nar_audio_embeddings[j](
+                    prompts[..., j]
+                )
+
+            for i, (predict_layer, embedding_layer) in enumerate(
+                zip(
+                    self.nar_predict_layers,
+                    self.nar_audio_embeddings[1:],
+                )
+            ):
+                y_pos = self.nar_audio_prenet(y_emb)
+                y_pos = self.nar_audio_position(y_pos)
+                xy_pos = torch.concat([x, y_pos], dim=1)
+
+                xy_dec, _ = self.nar_decoder(
+                    (xy_pos, self.nar_stage_embeddings[i].weight)
+                )
+                logits = predict_layer(xy_dec[:, text_len + prefix_len :])
+
+                samples = torch.argmax(logits, dim=-1)
+                codes.append(samples)
+
+                if i < 6:
+                    y_emb[:, prefix_len:] += embedding_layer(samples)
+
+        assert len(codes) == 8
+        return torch.stack(codes, dim=-1)
+
+
+# https://github.com/microsoft/unilm/blob/master/xtune/src/transformers/modeling_utils.py
+def top_k_top_p_filtering(
+    logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1
+):
+    """Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
+    Args:
+        logits: logits distribution shape (batch size, vocabulary size)
+        if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
+        if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
+            Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
+        Make sure we keep at least min_tokens_to_keep per batch example in the output
+    From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
+    """
+    if top_k > 0:
+        top_k = min(
+            max(top_k, min_tokens_to_keep), logits.size(-1)
+        )  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+
+    if top_p < 1.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(
+            F.softmax(sorted_logits, dim=-1), dim=-1
+        )
+
+        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > top_p
+        if min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., :min_tokens_to_keep] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[
+            ..., :-1
+        ].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(
+            1, sorted_indices, sorted_indices_to_remove
+        )
+        logits[indices_to_remove] = filter_value
+    return logits
+
+
+def topk_sampling(logits, top_k=10, top_p=1.0, temperature=1.0):
+    # temperature: (`optional`) float
+    #     The value used to module the next token probabilities. Must be strictly positive. Default to 1.0.
+    # top_k: (`optional`) int
+    #     The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.
+    # top_p: (`optional`) float
+    #     The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.
+
+    # Temperature (higher temperature => more likely to sample low probability tokens)
+    if temperature != 1.0:
+        logits = logits / temperature
+    # Top-p/top-k filtering
+    logits = top_k_top_p_filtering(logits, top_k=top_k, top_p=top_p)
+    # Sample
+    token = torch.multinomial(F.softmax(logits, dim=-1), num_samples=1)
+    logprobs = F.log_softmax(logits.float(), dim=-1)
+    current_logprobs = logprobs[torch.arange(logprobs.shape[0]), token.squeeze(1)]
+    return token, current_logprobs

models/visualizer.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# Copyright    2023                           (authors: Feiteng Li)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Dict, List, Tuple, Union
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+
+
+def visualize(
+    predicts: Tuple[torch.Tensor],
+    batch: Dict[str, Union[List, torch.Tensor]],
+    output_dir: str,
+    limit: int = 4,
+) -> None:
+    text_tokens = batch["text_tokens"].to("cpu").detach().numpy()
+    text_tokens_lens = batch["text_tokens_lens"].to("cpu").detach().numpy()
+    audio_features = batch["audio_features"].to("cpu").detach().numpy()
+    audio_features_lens = (
+        batch["audio_features_lens"].to("cpu").detach().numpy()
+    )
+    assert text_tokens.ndim == 2
+
+    utt_ids, texts = batch["utt_id"], batch["text"]
+
+    encoder_outputs = predicts[0].to("cpu").type(torch.float32).detach().numpy()
+    decoder_outputs = predicts[1]
+    if isinstance(decoder_outputs, list):
+        decoder_outputs = decoder_outputs[-1]
+    decoder_outputs = (
+        decoder_outputs.to("cpu").type(torch.float32).detach().numpy()
+    )
+
+    vmin, vmax = 0, 1024  # Encodec
+    if decoder_outputs.dtype == np.float32:
+        vmin, vmax = -6, 0  # Fbank
+
+    num_figures = 3
+    for b, (utt_id, text) in enumerate(zip(utt_ids[:limit], texts[:limit])):
+        _ = plt.figure(figsize=(14, 8 * num_figures))
+
+        S = text_tokens_lens[b]
+        T = audio_features_lens[b]
+
+        # encoder
+        plt.subplot(num_figures, 1, 1)
+        plt.title(f"Text: {text}")
+        plt.imshow(
+            X=np.transpose(encoder_outputs[b]),
+            cmap=plt.get_cmap("jet"),
+            aspect="auto",
+            interpolation="nearest",
+        )
+        plt.gca().invert_yaxis()
+        plt.axvline(x=S - 0.4, linewidth=2, color="r")
+        plt.xlabel("Encoder Output")
+        plt.colorbar()
+
+        # decoder
+        plt.subplot(num_figures, 1, 2)
+        plt.imshow(
+            X=np.transpose(decoder_outputs[b]),
+            cmap=plt.get_cmap("jet"),
+            aspect="auto",
+            interpolation="nearest",
+            vmin=vmin,
+            vmax=vmax,
+        )
+        plt.gca().invert_yaxis()
+        plt.axvline(x=T - 0.4, linewidth=2, color="r")
+        plt.xlabel("Decoder Output")
+        plt.colorbar()
+
+        # target
+        plt.subplot(num_figures, 1, 3)
+        plt.imshow(
+            X=np.transpose(audio_features[b]),
+            cmap=plt.get_cmap("jet"),
+            aspect="auto",
+            interpolation="nearest",
+            vmin=vmin,
+            vmax=vmax,
+        )
+        plt.gca().invert_yaxis()
+        plt.axvline(x=T - 0.4, linewidth=2, color="r")
+        plt.xlabel("Decoder Target")
+        plt.colorbar()
+
+        plt.savefig(f"{output_dir}/{utt_id}.png")
+        plt.close()

modules/activation.py

@@ -0,0 +1,612 @@
+from typing import Optional, Tuple, List
+import math
+
+import torch
+from torch import Tensor
+from torch.nn import Linear, Module
+from torch.nn import functional as F
+from torch.nn.init import constant_, xavier_normal_, xavier_uniform_
+from torch.nn.modules.linear import NonDynamicallyQuantizableLinear
+from torch.nn.parameter import Parameter
+
+def _in_projection_packed(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    w: Tensor,
+    b: Optional[Tensor] = None,
+) -> List[Tensor]:
+    r"""
+    Performs the in-projection step of the attention operation, using packed weights.
+    Output is a triple containing projection tensors for query, key and value.
+
+    Args:
+        q, k, v: query, key and value tensors to be projected. For self-attention,
+            these are typically the same tensor; for encoder-decoder attention,
+            k and v are typically the same tensor. (We take advantage of these
+            identities for performance if they are present.) Regardless, q, k and v
+            must share a common embedding dimension; otherwise their shapes may vary.
+        w: projection weights for q, k and v, packed into a single tensor. Weights
+            are packed along dimension 0, in q, k, v order.
+        b: optional projection biases for q, k and v, packed into a single tensor
+            in q, k, v order.
+
+    Shape:
+        Inputs:
+        - q: :math:`(..., E)` where E is the embedding dimension
+        - k: :math:`(..., E)` where E is the embedding dimension
+        - v: :math:`(..., E)` where E is the embedding dimension
+        - w: :math:`(E * 3, E)` where E is the embedding dimension
+        - b: :math:`E * 3` where E is the embedding dimension
+
+        Output:
+        - in output list :math:`[q', k', v']`, each output tensor will have the
+            same shape as the corresponding input tensor.
+    """
+    E = q.size(-1)
+    if k is v:
+        if q is k:
+            # self-attention
+            return F.linear(q, w, b).chunk(3, dim=-1)
+        else:
+            # encoder-decoder attention
+            w_q, w_kv = w.split([E, E * 2])
+            if b is None:
+                b_q = b_kv = None
+            else:
+                b_q, b_kv = b.split([E, E * 2])
+            return (F.linear(q, w_q, b_q),) + F.linear(k, w_kv, b_kv).chunk(2, dim=-1)
+    else:
+        w_q, w_k, w_v = w.chunk(3)
+        if b is None:
+            b_q = b_k = b_v = None
+        else:
+            b_q, b_k, b_v = b.chunk(3)
+        return F.linear(q, w_q, b_q), F.linear(k, w_k, b_k), F.linear(v, w_v, b_v)
+
+def _scaled_dot_product_attention(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    attn_mask: Optional[Tensor] = None,
+    dropout_p: float = 0.0,
+) -> Tuple[Tensor, Tensor]:
+    r"""
+    Computes scaled dot product attention on query, key and value tensors, using
+    an optional attention mask if passed, and applying dropout if a probability
+    greater than 0.0 is specified.
+    Returns a tensor pair containing attended values and attention weights.
+
+    Args:
+        q, k, v: query, key and value tensors. See Shape section for shape details.
+        attn_mask: optional tensor containing mask values to be added to calculated
+            attention. May be 2D or 3D; see Shape section for details.
+        dropout_p: dropout probability. If greater than 0.0, dropout is applied.
+
+    Shape:
+        - q: :math:`(B, Nt, E)` where B is batch size, Nt is the target sequence length,
+            and E is embedding dimension.
+        - key: :math:`(B, Ns, E)` where B is batch size, Ns is the source sequence length,
+            and E is embedding dimension.
+        - value: :math:`(B, Ns, E)` where B is batch size, Ns is the source sequence length,
+            and E is embedding dimension.
+        - attn_mask: either a 3D tensor of shape :math:`(B, Nt, Ns)` or a 2D tensor of
+            shape :math:`(Nt, Ns)`.
+
+        - Output: attention values have shape :math:`(B, Nt, E)`; attention weights
+            have shape :math:`(B, Nt, Ns)`
+    """
+    B, Nt, E = q.shape
+    q = q / math.sqrt(E)
+    # (B, Nt, E) x (B, E, Ns) -> (B, Nt, Ns)
+    if attn_mask is not None:
+        attn = torch.baddbmm(attn_mask, q, k.transpose(-2, -1))
+    else:
+        attn = torch.bmm(q, k.transpose(-2, -1))
+
+    attn = F.softmax(attn, dim=-1)
+    if dropout_p > 0.0:
+        attn = F.dropout(attn, p=dropout_p)
+    # (B, Nt, Ns) x (B, Ns, E) -> (B, Nt, E)
+    output = torch.bmm(attn, v)
+    return output, attn
+
+def multi_head_attention_forward(
+        x,
+        ipw,
+        ipb,
+        opw,
+        opb,
+        n_head,
+        attn_mask,
+        past_kv=None,
+        use_cache=False,
+):
+    # x = x.transpose(1, 0)
+    # tgt_len, bsz, embed_dim = x.shape
+    # head_dim = embed_dim // n_head
+    # q, k, v = _in_projection_packed(x, x, x, ipw, ipb)
+    # q = q.contiguous().view(tgt_len, bsz * n_head, head_dim).transpose(0, 1)
+    # k = k.contiguous().view(k.shape[0], bsz * n_head, head_dim).transpose(0, 1)
+    # v = v.contiguous().view(v.shape[0], bsz * n_head, head_dim).transpose(0, 1)
+
+    # new_attn_mask = torch.zeros_like(attn_mask, dtype=q.dtype)
+    # new_attn_mask.masked_fill_(attn_mask, float("-inf"))
+    # attn_mask = new_attn_mask
+    #
+    # attn_output, attn_output_weights = _scaled_dot_product_attention(q, k, v, attn_mask, 0.0)
+    # attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len * bsz, embed_dim)
+    # attn_output = torch._C._nn.linear(attn_output, opw, opb)
+    # attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
+
+    B, T, C = x.size()
+
+    q, k, v = torch._C._nn.linear(x, ipw, ipb).chunk(3, dim=-1)
+    k = k.view(B, T, n_head, C // n_head).transpose(1, 2)  # (B, nh, T, hs)
+    q = q.view(B, T, n_head, C // n_head).transpose(1, 2)  # (B, nh, T, hs)
+    v = v.view(B, T, n_head, C // n_head).transpose(1, 2)  # (B, nh, T, hs)
+    if past_kv is not None:
+        past_key = past_kv[0]
+        past_value = past_kv[1]
+        k = torch.cat((past_key, k), dim=-2)
+        v = torch.cat((past_value, v), dim=-2)
+
+    FULL_T = k.shape[-2]
+
+    if use_cache is True:
+        present = (k, v)
+    else:
+        present = None
+
+    att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+    att = att.masked_fill(attn_mask[FULL_T - T:FULL_T, :FULL_T], float('-inf'))
+    att = F.softmax(att, dim=-1)
+    y = att @ v  # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+    y = y.transpose(1, 2).contiguous().view(B, T, C)  # re-assemble all head outputs side by side
+    y = torch._C._nn.linear(y, opw, opb)
+    return (y, present)
+
+
+class MultiheadAttention(Module):
+    r"""Allows the model to jointly attend to information
+    from different representation subspaces as described in the paper:
+    `Attention Is All You Need <https://arxiv.org/abs/1706.03762>`_.
+
+    Multi-Head Attention is defined as:
+
+    .. math::
+        \text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
+
+    where :math:`head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)`.
+
+    ``forward()`` will use a special optimized implementation if all of the following
+    conditions are met:
+
+    - self attention is being computed (i.e., ``query``, ``key``, and ``value`` are the same tensor. This
+      restriction will be loosened in the future.)
+    - Either autograd is disabled (using ``torch.inference_mode`` or ``torch.no_grad``) or no tensor argument ``requires_grad``
+    - training is disabled (using ``.eval()``)
+    - dropout is 0
+    - ``add_bias_kv`` is ``False``
+    - ``add_zero_attn`` is ``False``
+    - ``batch_first`` is ``True`` and the input is batched
+    - ``kdim`` and ``vdim`` are equal to ``embed_dim``
+    - at most one of ``key_padding_mask`` or ``attn_mask`` is passed
+    - if a `NestedTensor <https://pytorch.org/docs/stable/nested.html>`_ is passed, neither ``key_padding_mask``
+      nor ``attn_mask`` is passed
+
+    If the optimized implementation is in use, a
+    `NestedTensor <https://pytorch.org/docs/stable/nested.html>`_ can be passed for
+    ``query``/``key``/``value`` to represent padding more efficiently than using a
+    padding mask. In this case, a `NestedTensor <https://pytorch.org/docs/stable/nested.html>`_
+    will be returned, and an additional speedup proportional to the fraction of the input
+    that is padding can be expected.
+
+    Args:
+        embed_dim: Total dimension of the model.
+        num_heads: Number of parallel attention heads. Note that ``embed_dim`` will be split
+            across ``num_heads`` (i.e. each head will have dimension ``embed_dim // num_heads``).
+        dropout: Dropout probability on ``attn_output_weights``. Default: ``0.0`` (no dropout).
+        bias: If specified, adds bias to input / output projection layers. Default: ``True``.
+        add_bias_kv: If specified, adds bias to the key and value sequences at dim=0. Default: ``False``.
+        add_zero_attn: If specified, adds a new batch of zeros to the key and value sequences at dim=1.
+            Default: ``False``.
+        kdim: Total number of features for keys. Default: ``None`` (uses ``kdim=embed_dim``).
+        vdim: Total number of features for values. Default: ``None`` (uses ``vdim=embed_dim``).
+        batch_first: If ``True``, then the input and output tensors are provided
+            as (batch, seq, feature). Default: ``False`` (seq, batch, feature).
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> multihead_attn = nn.MultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value)
+
+    """
+    __constants__ = ["batch_first"]
+    bias_k: Optional[torch.Tensor]
+    bias_v: Optional[torch.Tensor]
+
+    def __init__(
+            self,
+            embed_dim,
+            num_heads,
+            dropout=0.0,
+            bias=True,
+            add_bias_kv=False,
+            add_zero_attn=False,
+            kdim=None,
+            vdim=None,
+            batch_first=False,
+            linear1_cls=Linear,
+            linear2_cls=Linear,
+            device=None,
+            dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(MultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self._qkv_same_embed_dim = (
+                self.kdim == embed_dim and self.vdim == embed_dim
+        )
+
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.batch_first = batch_first
+        self.head_dim = embed_dim // num_heads
+        assert (
+                self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        if add_bias_kv:
+            self.bias_k = Parameter(
+                torch.empty((1, 1, embed_dim), **factory_kwargs)
+            )
+            self.bias_v = Parameter(
+                torch.empty((1, 1, embed_dim), **factory_kwargs)
+            )
+        else:
+            self.bias_k = self.bias_v = None
+
+        if linear1_cls == Linear:
+            if not self._qkv_same_embed_dim:
+                self.q_proj_weight = Parameter(
+                    torch.empty((embed_dim, embed_dim), **factory_kwargs)
+                )
+                self.k_proj_weight = Parameter(
+                    torch.empty((embed_dim, self.kdim), **factory_kwargs)
+                )
+                self.v_proj_weight = Parameter(
+                    torch.empty((embed_dim, self.vdim), **factory_kwargs)
+                )
+                self.register_parameter("in_proj_weight", None)
+            else:
+                self.in_proj_weight = Parameter(
+                    torch.empty((3 * embed_dim, embed_dim), **factory_kwargs)
+                )
+                self.register_parameter("q_proj_weight", None)
+                self.register_parameter("k_proj_weight", None)
+                self.register_parameter("v_proj_weight", None)
+
+            if bias:
+                self.in_proj_bias = Parameter(
+                    torch.empty(3 * embed_dim, **factory_kwargs)
+                )
+            else:
+                self.register_parameter("in_proj_bias", None)
+            self.out_proj = NonDynamicallyQuantizableLinear(
+                embed_dim, embed_dim, bias=bias, **factory_kwargs
+            )
+
+            self._reset_parameters()
+        else:
+            if not self._qkv_same_embed_dim:
+                raise NotImplementedError
+            else:
+                self.in_proj_linear = linear1_cls(
+                    embed_dim, 3 * embed_dim, bias=bias, **factory_kwargs
+                )
+                self.in_proj_weight = self.in_proj_linear.weight
+
+                self.register_parameter("q_proj_weight", None)
+                self.register_parameter("k_proj_weight", None)
+                self.register_parameter("v_proj_weight", None)
+
+                if bias:
+                    self.in_proj_bias = self.in_proj_linear.bias
+                else:
+                    self.register_parameter("in_proj_bias", None)
+
+            self.out_proj = linear2_cls(
+                embed_dim, embed_dim, bias=bias, **factory_kwargs
+            )
+
+            if self.bias_k is not None:
+                xavier_normal_(self.bias_k)
+            if self.bias_v is not None:
+                xavier_normal_(self.bias_v)
+
+        self.add_zero_attn = add_zero_attn
+
+    def _reset_parameters(self):
+        if self._qkv_same_embed_dim:
+            xavier_uniform_(self.in_proj_weight)
+        else:
+            xavier_uniform_(self.q_proj_weight)
+            xavier_uniform_(self.k_proj_weight)
+            xavier_uniform_(self.v_proj_weight)
+
+        if self.in_proj_bias is not None:
+            constant_(self.in_proj_bias, 0.0)
+            constant_(self.out_proj.bias, 0.0)
+
+        if self.bias_k is not None:
+            xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            xavier_normal_(self.bias_v)
+
+    def __setstate__(self, state):
+        # Support loading old MultiheadAttention checkpoints generated by v1.1.0
+        if "_qkv_same_embed_dim" not in state:
+            state["_qkv_same_embed_dim"] = True
+
+        super(MultiheadAttention, self).__setstate__(state)
+
+    def forward(
+            self,
+            query: Tensor,
+            key: Tensor,
+            value: Tensor,
+            key_padding_mask: Optional[Tensor] = None,
+            need_weights: bool = True,
+            attn_mask: Optional[Tensor] = None,
+            average_attn_weights: bool = True,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query: Query embeddings of shape :math:`(L, E_q)` for unbatched input, :math:`(L, N, E_q)` when ``batch_first=False``
+                or :math:`(N, L, E_q)` when ``batch_first=True``, where :math:`L` is the target sequence length,
+                :math:`N` is the batch size, and :math:`E_q` is the query embedding dimension ``embed_dim``.
+                Queries are compared against key-value pairs to produce the output.
+                See "Attention Is All You Need" for more details.
+            key: Key embeddings of shape :math:`(S, E_k)` for unbatched input, :math:`(S, N, E_k)` when ``batch_first=False``
+                or :math:`(N, S, E_k)` when ``batch_first=True``, where :math:`S` is the source sequence length,
+                :math:`N` is the batch size, and :math:`E_k` is the key embedding dimension ``kdim``.
+                See "Attention Is All You Need" for more details.
+            value: Value embeddings of shape :math:`(S, E_v)` for unbatched input, :math:`(S, N, E_v)` when
+                ``batch_first=False`` or :math:`(N, S, E_v)` when ``batch_first=True``, where :math:`S` is the source
+                sequence length, :math:`N` is the batch size, and :math:`E_v` is the value embedding dimension ``vdim``.
+                See "Attention Is All You Need" for more details.
+            key_padding_mask: If specified, a mask of shape :math:`(N, S)` indicating which elements within ``key``
+                to ignore for the purpose of attention (i.e. treat as "padding"). For unbatched `query`, shape should be :math:`(S)`.
+                Binary and byte masks are supported.
+                For a binary mask, a ``True`` value indicates that the corresponding ``key`` value will be ignored for
+                the purpose of attention. For a float mask, it will be directly added to the corresponding ``key`` value.
+            need_weights: If specified, returns ``attn_output_weights`` in addition to ``attn_outputs``.
+                Default: ``True``.
+            attn_mask: If specified, a 2D or 3D mask preventing attention to certain positions. Must be of shape
+                :math:`(L, S)` or :math:`(N\cdot\text{num\_heads}, L, S)`, where :math:`N` is the batch size,
+                :math:`L` is the target sequence length, and :math:`S` is the source sequence length. A 2D mask will be
+                broadcasted across the batch while a 3D mask allows for a different mask for each entry in the batch.
+                Binary, byte, and float masks are supported. For a binary mask, a ``True`` value indicates that the
+                corresponding position is not allowed to attend. For a byte mask, a non-zero value indicates that the
+                corresponding position is not allowed to attend. For a float mask, the mask values will be added to
+                the attention weight.
+            average_attn_weights: If true, indicates that the returned ``attn_weights`` should be averaged across
+                heads. Otherwise, ``attn_weights`` are provided separately per head. Note that this flag only has an
+                effect when ``need_weights=True``. Default: ``True`` (i.e. average weights across heads)
+
+        Outputs:
+            - **attn_output** - Attention outputs of shape :math:`(L, E)` when input is unbatched,
+              :math:`(L, N, E)` when ``batch_first=False`` or :math:`(N, L, E)` when ``batch_first=True``,
+              where :math:`L` is the target sequence length, :math:`N` is the batch size, and :math:`E` is the
+              embedding dimension ``embed_dim``.
+            - **attn_output_weights** - Only returned when ``need_weights=True``. If ``average_attn_weights=True``,
+              returns attention weights averaged across heads of shape :math:`(L, S)` when input is unbatched or
+              :math:`(N, L, S)`, where :math:`N` is the batch size, :math:`L` is the target sequence length, and
+              :math:`S` is the source sequence length. If ``average_attn_weights=False``, returns attention weights per
+              head of shape :math:`(\text{num\_heads}, L, S)` when input is unbatched or :math:`(N, \text{num\_heads}, L, S)`.
+
+            .. note::
+                `batch_first` argument is ignored for unbatched inputs.
+        """
+        is_batched = query.dim() == 3
+        if key_padding_mask is not None:
+            _kpm_dtype = key_padding_mask.dtype
+            if _kpm_dtype != torch.bool and not torch.is_floating_point(
+                    key_padding_mask
+            ):
+                raise AssertionError(
+                    "only bool and floating types of key_padding_mask are supported"
+                )
+        why_not_fast_path = ""
+        if not is_batched:
+            why_not_fast_path = f"input not batched; expected query.dim() of 3 but got {query.dim()}"
+        elif query is not key or key is not value:
+            # When lifting this restriction, don't forget to either
+            # enforce that the dtypes all match or test cases where
+            # they don't!
+            why_not_fast_path = "non-self attention was used (query, key, and value are not the same Tensor)"
+        elif (
+                self.in_proj_bias is not None
+                and query.dtype != self.in_proj_bias.dtype
+        ):
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_bias ({self.in_proj_bias.dtype}) don't match"
+        elif (
+                self.in_proj_weight is not None
+                and query.dtype != self.in_proj_weight.dtype
+        ):
+            # this case will fail anyway, but at least they'll get a useful error message.
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_weight ({self.in_proj_weight.dtype}) don't match"
+        elif self.training:
+            why_not_fast_path = "training is enabled"
+        elif not self.batch_first:
+            why_not_fast_path = "batch_first was not True"
+        elif self.bias_k is not None:
+            why_not_fast_path = "self.bias_k was not None"
+        elif self.bias_v is not None:
+            why_not_fast_path = "self.bias_v was not None"
+        elif self.dropout:
+            why_not_fast_path = f"dropout was {self.dropout}, required zero"
+        elif self.add_zero_attn:
+            why_not_fast_path = "add_zero_attn was enabled"
+        elif not self._qkv_same_embed_dim:
+            why_not_fast_path = "_qkv_same_embed_dim was not True"
+        elif attn_mask is not None:
+            why_not_fast_path = "attn_mask was not None"
+        elif query.is_nested and key_padding_mask is not None:
+            why_not_fast_path = (
+                "key_padding_mask is not supported with NestedTensor input"
+            )
+        elif self.num_heads % 2 == 1:
+            why_not_fast_path = "num_heads is odd"
+        elif torch.is_autocast_enabled():
+            why_not_fast_path = "autocast is enabled"
+
+        if not why_not_fast_path:
+            tensor_args = (
+                query,
+                key,
+                value,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.out_proj.weight,
+                self.out_proj.bias,
+            )
+            # We have to use list comprehensions below because TorchScript does not support
+            # generator expressions.
+            if torch.overrides.has_torch_function(tensor_args):
+                why_not_fast_path = "some Tensor argument has_torch_function"
+            elif not all(
+                    [
+                        (x is None or x.is_cuda or "cpu" in str(x.device))
+                        for x in tensor_args
+                    ]
+            ):
+                why_not_fast_path = (
+                    "some Tensor argument is neither CUDA nor CPU"
+                )
+            elif torch.is_grad_enabled() and any(
+                    [x is not None and x.requires_grad for x in tensor_args]
+            ):
+                why_not_fast_path = (
+                    "grad is enabled and at least one of query or the "
+                    "input/output projection weights or biases requires_grad"
+                )
+            if not why_not_fast_path:
+                return torch._native_multi_head_attention(
+                    query,
+                    key,
+                    value,
+                    self.embed_dim,
+                    self.num_heads,
+                    self.in_proj_weight,
+                    self.in_proj_bias,
+                    self.out_proj.weight,
+                    self.out_proj.bias,
+                    key_padding_mask
+                    if key_padding_mask is not None
+                    else attn_mask,
+                    need_weights,
+                    average_attn_weights,
+                    1
+                    if key_padding_mask is not None
+                    else 0
+                    if attn_mask is not None
+                    else None,
+                )
+
+        any_nested = query.is_nested or key.is_nested or value.is_nested
+        assert not any_nested, (
+                "MultiheadAttention does not support NestedTensor outside of its fast path. "
+                + f"The fast path was not hit because {why_not_fast_path}"
+        )
+
+        if self.batch_first and is_batched:
+            # make sure that the transpose op does not affect the "is" property
+            if key is value:
+                if query is key:
+                    query = key = value = query.transpose(1, 0)
+                else:
+                    query, key = [x.transpose(1, 0) for x in (query, key)]
+                    value = key
+            else:
+                query, key, value = [
+                    x.transpose(1, 0) for x in (query, key, value)
+                ]
+
+        if not self._qkv_same_embed_dim:
+            attn_output, attn_output_weights = F.multi_head_attention_forward(
+                query,
+                key,
+                value,
+                self.embed_dim,
+                self.num_heads,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.bias_k,
+                self.bias_v,
+                self.add_zero_attn,
+                self.dropout,
+                self.out_proj.weight,
+                self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                use_separate_proj_weight=True,
+                q_proj_weight=self.q_proj_weight,
+                k_proj_weight=self.k_proj_weight,
+                v_proj_weight=self.v_proj_weight,
+                average_attn_weights=average_attn_weights,
+            )
+        else:
+            attn_output, attn_output_weights = F.multi_head_attention_forward(
+                query,
+                key,
+                value,
+                self.embed_dim,
+                self.num_heads,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.bias_k,
+                self.bias_v,
+                self.add_zero_attn,
+                self.dropout,
+                self.out_proj.weight,
+                self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                average_attn_weights=average_attn_weights,
+            )
+        if self.batch_first and is_batched:
+            return attn_output.transpose(1, 0), attn_output_weights
+        else:
+            return attn_output, attn_output_weights
+
+    def infer(self,
+              x: Tensor,
+              key_padding_mask: Optional[Tensor] = None,
+              need_weights: bool = True,
+              attn_mask: Optional[Tensor] = None,
+              average_attn_weights: bool = True,
+              past_kv = None,
+              use_cache = False
+              ):
+        # x = x.transpose(1, 0)
+        y, kv = multi_head_attention_forward(
+                x=x,
+                ipw=self.in_proj_weight,
+                ipb=self.in_proj_bias,
+                opw=self.out_proj.weight,
+                opb=self.out_proj.bias,
+                n_head=self.num_heads,
+                attn_mask=attn_mask,
+                past_kv=past_kv,
+                use_cache=use_cache,
+        )
+        return (y, kv)

modules/embedding.py

@@ -0,0 +1,97 @@
+# Copyright    2023                             (authors: Feiteng Li)
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+
+import torch
+import torch.nn as nn
+
+
+class TokenEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim_model: int,
+        vocab_size: int,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+
+        self.vocab_size = vocab_size
+        self.dim_model = dim_model
+
+        self.dropout = torch.nn.Dropout(p=dropout)
+        self.word_embeddings = nn.Embedding(self.vocab_size, self.dim_model)
+
+    @property
+    def weight(self) -> torch.Tensor:
+        return self.word_embeddings.weight
+
+    def embedding(self, index: int) -> torch.Tensor:
+        return self.word_embeddings.weight[index : index + 1]
+
+    def forward(self, x: torch.Tensor):
+        X = self.word_embeddings(x)
+        X = self.dropout(X)
+
+        return X
+
+
+class SinePositionalEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim_model: int,
+        dropout: float = 0.0,
+        scale: bool = False,
+        alpha: bool = False,
+    ):
+        super().__init__()
+        self.dim_model = dim_model
+        self.x_scale = math.sqrt(dim_model) if scale else 1.0
+        self.alpha = nn.Parameter(torch.ones(1), requires_grad=alpha)
+        self.dropout = torch.nn.Dropout(p=dropout)
+
+        self.reverse = False
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, 4000))
+
+    def extend_pe(self, x):
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.dim_model)
+        if self.reverse:
+            position = torch.arange(
+                x.size(1) - 1, -1, -1.0, dtype=torch.float32
+            ).unsqueeze(1)
+        else:
+            position = torch.arange(
+                0, x.size(1), dtype=torch.float32
+            ).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.dim_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.dim_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        self.pe = pe.to(device=x.device, dtype=x.dtype).detach()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        self.extend_pe(x)
+        output = x.unsqueeze(-1) if x.ndim == 2 else x
+        output = output * self.x_scale + self.alpha * self.pe[:, : x.size(1)]
+        return self.dropout(output)

modules/optim.py

@@ -0,0 +1,1105 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import contextlib
+import logging
+import random
+from collections import defaultdict
+from typing import List, Optional, Tuple, Union
+
+import torch
+from lhotse.utils import fix_random_seed
+from torch import Tensor
+from torch.optim import Optimizer
+
+
+class BatchedOptimizer(Optimizer):
+    """
+    This class adds to class Optimizer the capability to optimize parameters in batches:
+    it will stack the parameters and their grads for you so the optimizer can work
+    on tensors with an extra leading dimension.  This is intended for speed with GPUs,
+    as it reduces the number of kernels launched in the optimizer.
+
+    Args:
+      params:
+    """
+
+    def __init__(self, params, defaults):
+        super(BatchedOptimizer, self).__init__(params, defaults)
+
+    @contextlib.contextmanager
+    def batched_params(self, param_group, group_params_names):
+        """
+        This function returns (technically, yields) a list of
+          of tuples (p, state), where
+        p is a `fake` parameter that is stacked (over axis 0) from real parameters
+        that share the same shape, and its gradient is also stacked;
+        `state` is the state corresponding to this batch of parameters
+        (it will be physically located in the "state" for one of the real
+        parameters, the last one that has any particular shape and dtype).
+
+        This function is decorated as a context manager so that it can
+        write parameters back to their "real" locations.
+
+        The idea is, instead of doing:
+        <code>
+          for p in group["params"]:
+             state = self.state[p]
+             ...
+        </code>
+        you can do:
+        <code>
+          with self.batched_params(group["params"]) as batches:
+             for p, state, p_names in batches:
+                 ...
+        </code>
+
+        Args:
+          group: a parameter group, which is a list of parameters; should be
+                one of self.param_groups.
+          group_params_names: name for each parameter in group,
+                which is List[str].
+        """
+        batches = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
+        batches_names = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of str
+
+        assert len(param_group) == len(group_params_names)
+        for p, named_p in zip(param_group, group_params_names):
+            key = (str(p.dtype), *p.shape)
+            batches[key].append(p)
+            batches_names[key].append(named_p)
+
+        batches_names_keys = list(batches_names.keys())
+        sorted_idx = sorted(
+            range(len(batches_names)), key=lambda i: batches_names_keys[i]
+        )
+        batches_names = [
+            batches_names[batches_names_keys[idx]] for idx in sorted_idx
+        ]
+        batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
+
+        stacked_params_dict = dict()
+
+        # turn batches into a list, in deterministic order.
+        # tuples will contain tuples of (stacked_param, state, stacked_params_names),
+        # one for each batch in `batches`.
+        tuples = []
+
+        for batch, batch_names in zip(batches, batches_names):
+            p = batch[0]
+            # we arbitrarily store the state in the
+            # state corresponding to the 1st parameter in the
+            # group.  class Optimizer will take care of saving/loading state.
+            state = self.state[p]
+            p_stacked = torch.stack(batch)
+            grad = torch.stack(
+                [
+                    torch.zeros_like(p) if p.grad is None else p.grad
+                    for p in batch
+                ]
+            )
+            p_stacked.grad = grad
+            stacked_params_dict[key] = p_stacked
+            tuples.append((p_stacked, state, batch_names))
+
+        yield tuples  # <-- calling code will do the actual optimization here!
+
+        for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
+            for i, p in enumerate(batch):  # batch is list of Parameter
+                p.copy_(stacked_params[i])
+
+
+class ScaledAdam(BatchedOptimizer):
+    """
+     Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update
+     proportional to the norm of that parameter; and also learn the scale of the parameter,
+     in log space, subject to upper and lower limits (as if we had factored each parameter as
+     param = underlying_param * log_scale.exp())
+
+
+     Args:
+          params:  The parameters or param_groups to optimize (like other Optimizer subclasses)
+              lr:  The learning rate.  We will typically use a learning rate schedule that starts
+                   at 0.03 and decreases over time, i.e. much higher than other common
+                   optimizers.
+     clipping_scale: (e.g. 2.0)
+                   A scale for gradient-clipping: if specified, the normalized gradients
+                   over the whole model will be clipped to have 2-norm equal to
+                   `clipping_scale` times the median 2-norm over the most recent period
+                   of `clipping_update_period` minibatches.  By "normalized gradients",
+                   we mean after multiplying by the rms parameter value for this tensor
+                   [for non-scalars]; this is appropriate because our update is scaled
+                   by this quantity.
+            betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad.
+                   Must satisfy 0 < beta <= beta2 < 1.
+     scalar_lr_scale: A scaling factor on the learning rate, that we use to update the
+                   scale of each parameter tensor and scalar parameters of the mode..
+                   If each parameter were decomposed
+                   as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale
+                   would be a the scaling factor on the learning rate of p_scale.
+              eps:  A general-purpose epsilon to prevent division by zero
+    param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be >= this value)
+    param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be <= this value)
+       scalar_max: Maximum absolute value for scalar parameters (applicable if your
+                   model has any parameters with numel() == 1).
+    size_update_period: The periodicity, in steps, with which we update the size (scale)
+                   of the parameter tensor.  This is provided to save a little time
+                   in the update.
+     clipping_update_period: if clipping_scale is specified, this is the period
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=3e-02,
+        clipping_scale=None,
+        betas=(0.9, 0.98),
+        scalar_lr_scale=0.1,
+        eps=1.0e-08,
+        param_min_rms=1.0e-05,
+        param_max_rms=3.0,
+        scalar_max=10.0,
+        size_update_period=4,
+        clipping_update_period=100,
+        parameters_names=None,
+        show_dominant_parameters=True,
+    ):
+
+        assert parameters_names is not None, (
+            "Please prepare parameters_names,"
+            "which is a List[List[str]]. Each List[str] is for a group"
+            "and each str is for a parameter"
+        )
+        defaults = dict(
+            lr=lr,
+            clipping_scale=clipping_scale,
+            betas=betas,
+            scalar_lr_scale=scalar_lr_scale,
+            eps=eps,
+            param_min_rms=param_min_rms,
+            param_max_rms=param_max_rms,
+            scalar_max=scalar_max,
+            size_update_period=size_update_period,
+            clipping_update_period=clipping_update_period,
+        )
+
+        super(ScaledAdam, self).__init__(params, defaults)
+        assert len(self.param_groups) == len(parameters_names)
+        self.parameters_names = parameters_names
+        self.show_dominant_parameters = show_dominant_parameters
+
+    def __setstate__(self, state):
+        super(ScaledAdam, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        batch = True
+
+        for group, group_params_names in zip(
+            self.param_groups, self.parameters_names
+        ):
+
+            with self.batched_params(
+                group["params"], group_params_names
+            ) as batches:
+
+                # batches is list of pairs (stacked_param, state).  stacked_param is like
+                # a regular parameter, and will have a .grad, but the 1st dim corresponds to
+                # a stacking dim, it is not a real dim.
+
+                if (
+                    len(batches[0][1]) == 0
+                ):  # if len(first state) == 0: not yet initialized
+                    clipping_scale = 1
+                else:
+                    clipping_scale = self._get_clipping_scale(group, batches)
+
+                for p, state, _ in batches:
+                    # Perform optimization step.
+                    # grad is not going to be None, we handled that when creating the batches.
+                    grad = p.grad
+                    if grad.is_sparse:
+                        raise RuntimeError(
+                            "ScaledAdam optimizer does not support sparse gradients"
+                        )
+                    # State initialization
+                    if len(state) == 0:
+                        self._init_state(group, p, state)
+
+                    self._step_one_batch(group, p, state, clipping_scale)
+
+        return loss
+
+    def _init_state(self, group: dict, p: Tensor, state: dict):
+        """
+        Initializes state dict for parameter 'p'.  Assumes that dim 0 of tensor p
+        is actually the batch dimension, corresponding to batched-together
+        parameters of a given shape.
+
+
+        Args:
+           group:   Dict to look up configuration values.
+               p: The parameter that we are initializing the state for
+           state: Dict from string to whatever state we are initializing
+        """
+        size_update_period = group["size_update_period"]
+
+        state["step"] = 0
+
+        kwargs = {"device": p.device, "dtype": p.dtype}
+
+        # 'delta' implements conventional momentum.  There are
+        # several different kinds of update going on, so rather than
+        # compute "exp_avg" like in Adam, we store and decay a
+        # parameter-change "delta", which combines all forms of
+        # update.  this is equivalent to how it's done in Adam,
+        # except for the first few steps.
+        state["delta"] = torch.zeros_like(
+            p, memory_format=torch.preserve_format
+        )
+
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        numel = p.numel()
+
+        if numel > 1:
+            # "param_rms" just periodically records the scalar root-mean-square value of
+            # the parameter tensor.
+            # it has a shape like (batch_size, 1, 1, 1, 1)
+            param_rms = (
+                (p ** 2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
+            )
+            state["param_rms"] = param_rms
+
+            state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
+            state["scale_grads"] = torch.zeros(
+                size_update_period, *param_rms.shape, **kwargs
+            )
+
+        # exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
+        state["exp_avg_sq"] = torch.zeros_like(
+            p, memory_format=torch.preserve_format
+        )
+
+    def _get_clipping_scale(
+        self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]]
+    ) -> float:
+        """
+        Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
+        by this amount before applying the rest of the update.
+
+        Args:
+           group: the parameter group, an item in self.param_groups
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+        """
+        assert len(tuples) >= 1
+        clipping_scale = group["clipping_scale"]
+        (first_p, first_state, _) = tuples[0]
+        step = first_state["step"]
+        if clipping_scale is None or step == 0:
+            # no clipping.  return early on step == 0 because the other
+            # parameters' state won't have been initialized yet.
+            return 1.0
+        clipping_update_period = group["clipping_update_period"]
+
+        tot_sumsq = torch.tensor(0.0, device=first_p.device)
+        for (p, state, param_names) in tuples:
+            grad = p.grad
+            if grad.is_sparse:
+                raise RuntimeError(
+                    "ScaledAdam optimizer does not support sparse gradients"
+                )
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                tot_sumsq += (
+                    grad ** 2
+                ).sum()  # sum() to change shape [1] to []
+            else:
+                tot_sumsq += ((grad * state["param_rms"]) ** 2).sum()
+
+        tot_norm = tot_sumsq.sqrt()
+        if "model_norms" not in first_state:
+            first_state["model_norms"] = torch.zeros(
+                clipping_update_period, device=p.device
+            )
+        first_state["model_norms"][step % clipping_update_period] = tot_norm
+
+        if step % clipping_update_period == 0:
+            # Print some stats.
+            # We don't reach here if step == 0 because we would have returned
+            # above.
+            sorted_norms = first_state["model_norms"].sort()[0].to("cpu")
+            quartiles = []
+            for n in range(0, 5):
+                index = min(
+                    clipping_update_period - 1,
+                    (clipping_update_period // 4) * n,
+                )
+                quartiles.append(sorted_norms[index].item())
+
+            median = quartiles[2]
+            threshold = clipping_scale * median
+            first_state["model_norm_threshold"] = threshold
+            percent_clipped = (
+                first_state["num_clipped"] * 100.0 / clipping_update_period
+                if "num_clipped" in first_state
+                else 0.0
+            )
+            first_state["num_clipped"] = 0
+            quartiles = " ".join(["%.3e" % x for x in quartiles])
+            logging.info(
+                f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
+                f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
+            )
+
+        if step < clipping_update_period:
+            return 1.0  # We have not yet estimated a norm to clip to.
+        else:
+            try:
+                model_norm_threshold = first_state["model_norm_threshold"]
+            except KeyError:
+                logging.info(
+                    "Warning: model_norm_threshold not in state: possibly "
+                    "you changed config when restarting, adding clipping_scale option?"
+                )
+                return 1.0
+            ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
+            if ans < 1.0:
+                first_state["num_clipped"] += 1
+            if ans < 0.1:
+                logging.warn(
+                    f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
+                )
+                if self.show_dominant_parameters:
+                    assert p.shape[0] == len(param_names)
+                    self._show_gradient_dominating_parameter(tuples, tot_sumsq)
+            return ans
+
+    def _show_gradient_dominating_parameter(
+        self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor
+    ):
+        """
+        Show information of parameter wihch dominanting tot_sumsq.
+
+        Args:
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+            tot_sumsq: sumsq of all parameters. Though it's could be calculated
+                from tuples, we still pass it to save some time.
+        """
+        all_sumsq_orig = {}
+        for (p, state, batch_param_names) in tuples:
+            # p is a stacked batch parameters.
+            batch_grad = p.grad
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                batch_sumsq_orig = batch_grad ** 2
+                # Dummpy values used by following `zip` statement.
+                batch_rms_orig = torch.ones(p.shape[0])
+            else:
+                batch_rms_orig = state["param_rms"]
+                batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum(
+                    dim=list(range(1, batch_grad.ndim))
+                )
+
+            for name, sumsq_orig, rms, grad in zip(
+                batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad
+            ):
+
+                proportion_orig = sumsq_orig / tot_sumsq
+                all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
+
+        assert torch.isclose(
+            sum([value[0] for value in all_sumsq_orig.values()]).cpu(),
+            torch.tensor(1.0),
+        )
+        sorted_by_proportion = {
+            k: v
+            for k, v in sorted(
+                all_sumsq_orig.items(),
+                key=lambda item: item[1][0],
+                reverse=True,
+            )
+        }
+        dominant_param_name = next(iter(sorted_by_proportion))
+        (
+            dominant_proportion,
+            dominant_sumsq,
+            dominant_rms,
+            dominant_grad,
+        ) = sorted_by_proportion[dominant_param_name]
+        logging.info(
+            f"Parameter Dominanting tot_sumsq {dominant_param_name}"
+            f" with proportion {dominant_proportion:.2f},"
+            f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
+            f"={dominant_sumsq:.3e},"
+            f" grad_sumsq = {(dominant_grad**2).sum():.3e},"
+            f" orig_rms_sq={(dominant_rms**2).item():.3e}"
+        )
+
+    def _step_one_batch(
+        self, group: dict, p: Tensor, state: dict, clipping_scale: float
+    ):
+        """
+        Do the step for one parameter, which is actually going to be a batch of
+        `real` parameters, with dim 0 as the batch dim.
+        Args:
+                  group:  dict to look up configuration values
+                    p: parameter to update (actually multiple parameters stacked together
+                       as a batch)
+                  state: state-dict for p, to look up the optimizer state
+        """
+        lr = group["lr"]
+        size_update_period = group["size_update_period"]
+        beta1 = group["betas"][0]
+
+        grad = p.grad
+        if clipping_scale != 1.0:
+            grad = grad * clipping_scale
+        step = state["step"]
+        delta = state["delta"]
+
+        delta.mul_(beta1)
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        if numel > 1:
+            # Update the size/scale of p, and set param_rms
+            scale_grads = state["scale_grads"]
+            scale_grads[step % size_update_period] = (p * grad).sum(
+                dim=list(range(1, p.ndim)), keepdim=True
+            )
+            if step % size_update_period == size_update_period - 1:
+                param_rms = state["param_rms"]  # shape: (batch_size, 1, 1, ..)
+                param_rms.copy_(
+                    (p ** 2)
+                    .mean(dim=list(range(1, p.ndim)), keepdim=True)
+                    .sqrt()
+                )
+                if step > 0:
+                    # self._size_update() learns the overall scale on the
+                    # parameter, by shrinking or expanding it.
+                    self._size_update(group, scale_grads, p, state)
+
+        if numel == 1:
+            # For parameters with 1 element we just use regular Adam.
+            # Updates delta.
+            self._step_scalar(group, p, state)
+        else:
+            self._step(group, p, state)
+
+        state["step"] = step + 1
+
+    def _size_update(
+        self, group: dict, scale_grads: Tensor, p: Tensor, state: dict
+    ) -> None:
+        """
+               Called only where p.numel() > 1, this updates the scale of the parameter.
+               If we imagine: p =  underlying_param * scale.exp(), and we are doing
+               gradient descent on underlying param and on scale, this function does the update
+               on `scale`.
+
+               Args:
+              group: dict to look up configuration values
+        scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing
+                      grads w.r.t. the scales.
+                  p:  The parameter to update
+               state: The state-dict of p
+        """
+
+        param_rms = state["param_rms"]
+        beta1, beta2 = group["betas"]
+        size_lr = group["lr"] * group["scalar_lr_scale"]
+        param_min_rms = group["param_min_rms"]
+        param_max_rms = group["param_max_rms"]
+        eps = group["eps"]
+        step = state["step"]
+        batch_size = p.shape[0]
+
+        size_update_period = scale_grads.shape[0]
+        # correct beta2 for the size update period: we will have
+        # faster decay at this level.
+        beta2_corr = beta2 ** size_update_period
+
+        scale_exp_avg_sq = state[
+            "scale_exp_avg_sq"
+        ]  # shape: (batch_size, 1, 1, ..)
+        scale_exp_avg_sq.mul_(beta2_corr).add_(
+            (scale_grads ** 2).mean(
+                dim=0
+            ),  # mean over dim `size_update_period`
+            alpha=1 - beta2_corr,
+        )  # shape is (batch_size, 1, 1, ...)
+
+        # The 1st time we reach here is when size_step == 1.
+        size_step = (step + 1) // size_update_period
+        bias_correction2 = 1 - beta2_corr ** size_step
+        # we don't bother with bias_correction1; this will help prevent divergence
+        # at the start of training.
+
+        denom = scale_exp_avg_sq.sqrt() + eps
+
+        scale_step = (
+            -size_lr
+            * (bias_correction2 ** 0.5)
+            * scale_grads.sum(dim=0)
+            / denom
+        )
+
+        is_too_small = param_rms < param_min_rms
+        is_too_large = param_rms > param_max_rms
+
+        # when the param gets too small, just don't shrink it any further.
+        scale_step.masked_fill_(is_too_small, 0.0)
+        # when it gets too large, stop it from getting any larger.
+        scale_step.masked_fill_(is_too_large, -size_lr * size_update_period)
+        delta = state["delta"]
+        # the factor of (1-beta1) relates to momentum.
+        delta.add_(p * scale_step, alpha=(1 - beta1))
+
+    def _step(self, group: dict, p: Tensor, state: dict):
+        """
+        This function does the core update of self.step(), in the case where the members of
+        the batch have more than 1 element.
+
+        Args:
+            group: A dict which will be used to look up configuration values
+                p: The parameter to be updated
+             grad: The grad of p
+            state: The state-dict corresponding to parameter p
+
+        This function modifies p.
+        """
+        grad = p.grad
+        lr = group["lr"]
+        beta1, beta2 = group["betas"]
+        eps = group["eps"]
+        param_min_rms = group["param_min_rms"]
+        step = state["step"]
+
+        exp_avg_sq = state["exp_avg_sq"]
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
+
+        this_step = state["step"] - (
+            state["zero_step"] if "zero_step" in state else 0
+        )
+        bias_correction2 = 1 - beta2 ** (this_step + 1)
+        if bias_correction2 < 0.99:
+            # note: not in-place.
+            exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
+
+        denom = exp_avg_sq.sqrt()
+        denom += eps
+        grad = grad / denom
+
+        alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms)
+
+        delta = state["delta"]
+        delta.add_(grad * alpha)
+        p.add_(delta)
+
+    def _step_scalar(self, group: dict, p: Tensor, state: dict):
+        """
+        A simplified form of the core update for scalar tensors, where we cannot get a good
+        estimate of the parameter rms.
+        """
+        beta1, beta2 = group["betas"]
+        scalar_max = group["scalar_max"]
+        eps = group["eps"]
+        lr = group["lr"] * group["scalar_lr_scale"]
+        grad = p.grad
+
+        exp_avg_sq = state["exp_avg_sq"]  # shape: (batch_size,)
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+        # bias_correction2 is like in Adam.  Don't bother with bias_correction1;
+        # slower update at the start will help stability anyway.
+        bias_correction2 = 1 - beta2 ** (state["step"] + 1)
+        denom = (exp_avg_sq / bias_correction2).sqrt() + eps
+
+        delta = state["delta"]
+        delta.add_(grad / denom, alpha=-lr * (1 - beta1))
+        p.clamp_(min=-scalar_max, max=scalar_max)
+        p.add_(delta)
+
+
+class LRScheduler(object):
+    """
+    Base-class for learning rate schedulers where the learning-rate depends on both the
+    batch and the epoch.
+    """
+
+    def __init__(self, optimizer: Optimizer, verbose: bool = False):
+        # Attach optimizer
+        if not isinstance(optimizer, Optimizer):
+            raise TypeError(
+                "{} is not an Optimizer".format(type(optimizer).__name__)
+            )
+        self.optimizer = optimizer
+        self.verbose = verbose
+
+        for group in optimizer.param_groups:
+            group.setdefault("base_lr", group["lr"])
+
+        self.base_lrs = [group["base_lr"] for group in optimizer.param_groups]
+
+        self.epoch = 0
+        self.batch = 0
+
+    def state_dict(self):
+        """Returns the state of the scheduler as a :class:`dict`.
+
+        It contains an entry for every variable in self.__dict__ which
+        is not the optimizer.
+        """
+        return {
+            "base_lrs": self.base_lrs,
+            "epoch": self.epoch,
+            "batch": self.batch,
+        }
+
+    def load_state_dict(self, state_dict):
+        """Loads the schedulers state.
+
+        Args:
+            state_dict (dict): scheduler state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self) -> List[float]:
+        """Return last computed learning rate by current scheduler.  Will be a list of float."""
+        return self._last_lr
+
+    def get_lr(self):
+        # Compute list of learning rates from self.epoch and self.batch and
+        # self.base_lrs; this must be overloaded by the user.
+        # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ]
+        raise NotImplementedError
+
+    def step_batch(self, batch: Optional[int] = None) -> None:
+        # Step the batch index, or just set it.  If `batch` is specified, it
+        # must be the batch index from the start of training, i.e. summed over
+        # all epochs.
+        # You can call this in any order; if you don't provide 'batch', it should
+        # of course be called once per batch.
+        if batch is not None:
+            self.batch = batch
+        else:
+            self.batch = self.batch + 1
+        self._set_lrs()
+
+    def step_epoch(self, epoch: Optional[int] = None):
+        # Step the epoch index, or just set it.  If you provide the 'epoch' arg,
+        # you should call this at the start of the epoch; if you don't provide the 'epoch'
+        # arg, you should call it at the end of the epoch.
+        if epoch is not None:
+            self.epoch = epoch
+        else:
+            self.epoch = self.epoch + 1
+        self._set_lrs()
+
+    def _set_lrs(self):
+        values = self.get_lr()
+        assert len(values) == len(self.optimizer.param_groups)
+
+        for i, data in enumerate(zip(self.optimizer.param_groups, values)):
+            param_group, lr = data
+            param_group["lr"] = lr
+            self.print_lr(self.verbose, i, lr)
+        self._last_lr = [group["lr"] for group in self.optimizer.param_groups]
+
+    def print_lr(self, is_verbose, group, lr):
+        """Display the current learning rate."""
+        if is_verbose:
+            logging.info(
+                f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate"
+                f" of group {group} to {lr:.4e}."
+            )
+
+
+class Eden(LRScheduler):
+    """
+    Eden scheduler.
+    The basic formula (before warmup) is:
+      lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 *
+                     (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup
+    where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches
+    and then stays constant at 1.
+
+
+     E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam
+
+    Args:
+        optimizer: the optimizer to change the learning rates on
+        lr_batches: the number of batches after which we start significantly
+              decreasing the learning rate, suggest 5000.
+        lr_epochs: the number of epochs after which we start significantly
+              decreasing the learning rate, suggest 6 if you plan to do e.g.
+              20 to 40 epochs, but may need smaller number if dataset is huge
+              and you will do few epochs.
+    """
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        lr_batches: Union[int, float],
+        lr_epochs: Union[int, float],
+        warmup_batches: Union[int, float] = 500.0,
+        verbose: bool = False,
+    ):
+        super(Eden, self).__init__(optimizer, verbose)
+        self.lr_batches = lr_batches
+        self.lr_epochs = lr_epochs
+        self.warmup_batches = warmup_batches
+
+    def get_lr(self):
+        factor = (
+            (self.batch ** 2 + self.lr_batches ** 2) / self.lr_batches ** 2
+        ) ** -0.25 * (
+            ((self.epoch ** 2 + self.lr_epochs ** 2) / self.lr_epochs ** 2)
+            ** -0.25
+        )
+        warmup_factor = (
+            1.0
+            if self.batch >= self.warmup_batches
+            else 0.5 + 0.5 * (self.batch / self.warmup_batches)
+        )
+
+        return [x * factor * warmup_factor for x in self.base_lrs]
+
+
+def _test_eden():
+    m = torch.nn.Linear(100, 100)
+    optim = ScaledAdam(m.parameters(), lr=0.03)
+
+    scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True)
+
+    for epoch in range(10):
+        scheduler.step_epoch(epoch)  # sets epoch to `epoch`
+
+        for step in range(20):
+            x = torch.randn(200, 100).detach()
+            x.requires_grad = True
+            y = m(x)
+            dy = torch.randn(200, 100).detach()
+            f = (y * dy).sum()
+            f.backward()
+
+            optim.step()
+            scheduler.step_batch()
+            optim.zero_grad()
+
+    logging.info(f"last lr = {scheduler.get_last_lr()}")
+    logging.info(f"state dict = {scheduler.state_dict()}")
+
+
+# This is included mostly as a baseline for ScaledAdam.
+class Eve(Optimizer):
+    """
+    Implements Eve algorithm.  This is a modified version of AdamW with a special
+    way of setting the weight-decay / shrinkage-factor, which is designed to make the
+    rms of the parameters approach a particular target_rms (default: 0.1).  This is
+    for use with networks with 'scaled' versions of modules (see scaling.py), which
+    will be close to invariant to the absolute scale on the parameter matrix.
+
+    The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_.
+    The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_.
+    Eve is unpublished so far.
+
+    Arguments:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float, optional): learning rate (default: 1e-3)
+        betas (Tuple[float, float], optional): coefficients used for computing
+            running averages of gradient and its square (default: (0.9, 0.999))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay coefficient (default: 3e-4;
+            this value means that the weight would decay significantly after
+            about 3k minibatches.  Is not multiplied by learning rate, but
+            is conditional on RMS-value of parameter being > target_rms.
+        target_rms (float, optional): target root-mean-square value of
+           parameters, if they fall below this we will stop applying weight decay.
+
+
+    .. _Adam: A Method for Stochastic Optimization:
+        https://arxiv.org/abs/1412.6980
+    .. _Decoupled Weight Decay Regularization:
+        https://arxiv.org/abs/1711.05101
+    .. _On the Convergence of Adam and Beyond:
+        https://openreview.net/forum?id=ryQu7f-RZ
+    """
+
+    def __init__(
+        self,
+        params,
+        lr=1e-3,
+        betas=(0.9, 0.98),
+        eps=1e-8,
+        weight_decay=1e-3,
+        target_rms=0.1,
+    ):
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError(
+                "Invalid beta parameter at index 0: {}".format(betas[0])
+            )
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError(
+                "Invalid beta parameter at index 1: {}".format(betas[1])
+            )
+        if not 0 <= weight_decay <= 0.1:
+            raise ValueError(
+                "Invalid weight_decay value: {}".format(weight_decay)
+            )
+        if not 0 < target_rms <= 10.0:
+            raise ValueError("Invalid target_rms value: {}".format(target_rms))
+        defaults = dict(
+            lr=lr,
+            betas=betas,
+            eps=eps,
+            weight_decay=weight_decay,
+            target_rms=target_rms,
+        )
+        super(Eve, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(Eve, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                # Perform optimization step
+                grad = p.grad
+                if grad.is_sparse:
+                    raise RuntimeError(
+                        "AdamW does not support sparse gradients"
+                    )
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state["step"] = 0
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(
+                        p, memory_format=torch.preserve_format
+                    )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+
+                # Decay the first and second moment running average coefficient
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+                denom = (exp_avg_sq.sqrt() * (bias_correction2 ** -0.5)).add_(
+                    group["eps"]
+                )
+
+                step_size = group["lr"] / bias_correction1
+                target_rms = group["target_rms"]
+                weight_decay = group["weight_decay"]
+
+                if p.numel() > 1:
+                    # avoid applying this weight-decay on "scaling factors"
+                    # (which are scalar).
+                    is_above_target_rms = p.norm() > (
+                        target_rms * (p.numel() ** 0.5)
+                    )
+                    p.mul_(1 - (weight_decay * is_above_target_rms))
+
+                p.addcdiv_(exp_avg, denom, value=-step_size)
+
+                # if random.random() < 0.0005:
+                #     step = (exp_avg / denom) * step_size
+                #     logging.info(
+                #         f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}"
+                #     )
+
+        return loss
+
+
+def _test_scaled_adam(hidden_dim: int):
+    import timeit
+
+    from scaling import ScaledLinear
+
+    E = 100
+    B = 4
+    T = 2
+    logging.info("in test_eve_cain")
+    # device = torch.device('cuda')
+    device = torch.device("cpu")
+    dtype = torch.float32
+
+    fix_random_seed(42)
+    # these input_magnitudes and output_magnitudes are to test that
+    # Abel is working as we expect and is able to adjust scales of
+    # different dims differently.
+    input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+    output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp()
+
+    for iter in [1, 0]:
+        fix_random_seed(42)
+        Linear = torch.nn.Linear if iter == 0 else ScaledLinear
+
+        m = torch.nn.Sequential(
+            Linear(E, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, hidden_dim),
+            torch.nn.PReLU(),
+            Linear(hidden_dim, E),
+        ).to(device)
+
+        train_pairs = [
+            (
+                100.0
+                * torch.randn(B, T, E, device=device, dtype=dtype)
+                * input_magnitudes,
+                torch.randn(B, T, E, device=device, dtype=dtype)
+                * output_magnitudes,
+            )
+            for _ in range(20)
+        ]
+
+        if iter == 0:
+            optim = Eve(m.parameters(), lr=0.003)
+        elif iter == 1:
+            optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0)
+        scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False)
+
+        start = timeit.default_timer()
+        avg_loss = 0.0
+        for epoch in range(180):
+            scheduler.step_epoch()
+            # if epoch == 100 and iter in [2,3]:
+            #    optim.reset_speedup()  # check it doesn't crash.
+
+            # if epoch == 130:
+            #    opts = diagnostics.TensorDiagnosticOptions(
+            #        2 ** 22
+            #    )  # allow 4 megabytes per sub-module
+            #    diagnostic = diagnostics.attach_diagnostics(m, opts)
+
+            for n, (x, y) in enumerate(train_pairs):
+                y_out = m(x)
+                loss = ((y_out - y) ** 2).mean() * 100.0
+                if epoch == 0 and n == 0:
+                    avg_loss = loss.item()
+                else:
+                    avg_loss = 0.98 * avg_loss + 0.02 * loss.item()
+                if n == 0 and epoch % 5 == 0:
+                    # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item()
+                    # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item()
+                    # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item()
+                    # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item()
+                    # scale1 = '%.2e' % (m[0].weight_scale.exp().item())
+                    # scale1b = '%.2e' % (m[0].bias_scale.exp().item())
+                    # scale2 = '%.2e' % (m[2].weight_scale.exp().item())
+                    # scale2b = '%.2e' % (m[2].bias_scale.exp().item())
+                    lr = scheduler.get_last_lr()[0]
+                    logging.info(
+                        f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}"
+                    )  # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b}
+                loss.log().backward()
+                optim.step()
+                optim.zero_grad()
+                scheduler.step_batch()
+
+        # diagnostic.print_diagnostics()
+
+        stop = timeit.default_timer()
+        logging.info(f"Iter={iter}, Time taken: {stop - start}")
+
+        logging.info(f"last lr = {scheduler.get_last_lr()}")
+        # logging.info("state dict = ", scheduler.state_dict())
+        # logging.info("optim state_dict = ", optim.state_dict())
+        logging.info(f"input_magnitudes = {input_magnitudes}")
+        logging.info(f"output_magnitudes = {output_magnitudes}")
+
+
+if __name__ == "__main__":
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    logging.getLogger().setLevel(logging.INFO)
+    import subprocess
+
+    s = subprocess.check_output(
+        "git status -uno .; git log -1; git diff HEAD .", shell=True
+    )
+    logging.info(s)
+    import sys
+
+    if len(sys.argv) > 1:
+        hidden_dim = int(sys.argv[1])
+    else:
+        hidden_dim = 200
+
+    _test_scaled_adam(hidden_dim)
+    _test_eden()

modules/scaling.py

@@ -0,0 +1,1401 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import collections
+import logging
+import random
+import math
+from functools import reduce
+from itertools import repeat
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from torch.nn import Embedding as ScaledEmbedding
+
+from utils import Transpose
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        scale_factor: Tensor,
+        sign_factor: Optional[Tensor],
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        xgt0 = x > 0
+        if sign_factor is None:
+            ctx.save_for_backward(xgt0, scale_factor)
+        else:
+            ctx.save_for_backward(xgt0, scale_factor, sign_factor)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]:
+        if len(ctx.saved_tensors) == 3:
+            xgt0, scale_factor, sign_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+                sign_factor = sign_factor.unsqueeze(-1)
+            factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        else:
+            xgt0, scale_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+            factor = scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        neg_delta_grad = x_grad.abs() * factor
+        return (
+            x_grad - neg_delta_grad,
+            None,
+            None,
+            None,
+        )
+
+
+def _compute_scale_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_abs: float,
+    max_abs: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    x_abs_mean = torch.mean(x.abs(), dim=sum_dims).to(torch.float32)
+
+    if min_abs == 0.0:
+        below_threshold = 0.0
+    else:
+        # below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if
+        # x_abs)_mean , min_abs.
+        below_threshold = (
+            (min_abs - x_abs_mean) * (gain_factor / min_abs)
+        ).clamp(min=0, max=max_factor)
+
+    above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp(
+        min=0, max=max_factor
+    )
+
+    return below_threshold - above_threshold
+
+
+def _compute_sign_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_positive: float,
+    max_positive: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    proportion_positive = torch.mean((x > 0).to(torch.float32), dim=sum_dims)
+    if min_positive == 0.0:
+        factor1 = 0.0
+    else:
+        # 0 if proportion_positive >= min_positive, else can be
+        # as large as max_factor.
+        factor1 = (
+            (min_positive - proportion_positive) * (gain_factor / min_positive)
+        ).clamp_(min=0, max=max_factor)
+
+    if max_positive == 1.0:
+        factor2 = 0.0
+    else:
+        # 0 if self.proportion_positive <= max_positive, else can be
+        # as large as -max_factor.
+        factor2 = (
+            (proportion_positive - max_positive)
+            * (gain_factor / (1.0 - max_positive))
+        ).clamp_(min=0, max=max_factor)
+    sign_factor = factor1 - factor2
+    # require min_positive != 0 or max_positive != 1:
+    assert not isinstance(sign_factor, float)
+    return sign_factor
+
+
+class ActivationScaleBalancerFunction(torch.autograd.Function):
+    """
+    This object is used in class ActivationBalancer when the user specified
+    min_positive=0, max_positive=1, so there are no constraints on the signs
+    of the activations and only the absolute value has a constraint.
+    """
+
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        sign_factor: Tensor,
+        scale_factor: Tensor,
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        xgt0 = x > 0
+        ctx.save_for_backward(xgt0, sign_factor, scale_factor)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]:
+        xgt0, sign_factor, scale_factor = ctx.saved_tensors
+        for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+            sign_factor = sign_factor.unsqueeze(-1)
+            scale_factor = scale_factor.unsqueeze(-1)
+
+        factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        neg_delta_grad = x_grad.abs() * factor
+        return (
+            x_grad - neg_delta_grad,
+            None,
+            None,
+            None,
+        )
+
+
+class RandomClampFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        min: Optional[float],
+        max: Optional[float],
+        prob: float,
+        reflect: float,
+    ) -> Tensor:
+        x_clamped = torch.clamp(x, min=min, max=max)
+        mask = torch.rand_like(x) < prob
+        ans = torch.where(mask, x_clamped, x)
+        if x.requires_grad:
+            ctx.save_for_backward(ans == x)
+            ctx.reflect = reflect
+        if reflect != 0.0:
+            ans = ans * (1.0 + reflect) - (x * reflect)
+        return ans
+
+    @staticmethod
+    def backward(
+        ctx, ans_grad: Tensor
+    ) -> Tuple[Tensor, None, None, None, None]:
+        (is_same,) = ctx.saved_tensors
+        x_grad = ans_grad * is_same.to(ans_grad.dtype)
+        reflect = ctx.reflect
+        if reflect != 0.0:
+            x_grad = x_grad * (1.0 + reflect) - (ans_grad * reflect)
+        return x_grad, None, None, None, None
+
+
+def random_clamp(
+    x: Tensor,
+    min: Optional[float] = None,
+    max: Optional[float] = None,
+    prob: float = 0.5,
+    reflect: float = 0.0,
+):
+    return RandomClampFunction.apply(x, min, max, prob, reflect)
+
+
+def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor:
+    """
+    A randomized way of casting a floating point value to half precision.
+    """
+    if x.dtype == torch.float16:
+        return x
+    x_abs = x.abs()
+    is_too_small = x_abs < min_abs
+    # for elements where is_too_small is true, random_val will contain +-min_abs with
+    # probability (x.abs() / min_abs), and 0.0 otherwise.  [so this preserves expectations,
+    # for those elements].
+    random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs)
+    return torch.where(is_too_small, random_val, x).to(torch.float16)
+
+
+class RandomGradFunction(torch.autograd.Function):
+    """
+    Does nothing in forward pass; in backward pass, gets rid of very small grads using
+    randomized approach that preserves expectations (intended to reduce roundoff).
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, min_abs: float) -> Tensor:
+        ctx.min_abs = min_abs
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None]:
+        if ans_grad.dtype == torch.float16:
+            return (
+                random_cast_to_half(
+                    ans_grad.to(torch.float32), min_abs=ctx.min_abs
+                ),
+                None,
+            )
+        else:
+            return ans_grad, None
+
+
+class RandomGrad(torch.nn.Module):
+    """
+    Gets rid of very small gradients using an expectation-preserving method, intended to increase
+    accuracy of training when using amp (automatic mixed precision)
+    """
+
+    def __init__(self, min_abs: float = 5.0e-06):
+        super(RandomGrad, self).__init__()
+        self.min_abs = min_abs
+
+    def forward(self, x: Tensor):
+        if (
+            torch.jit.is_scripting()
+            or not self.training
+            or torch.jit.is_tracing()
+        ):
+            return x
+        else:
+            return RandomGradFunction.apply(x, self.min_abs)
+
+
+class SoftmaxFunction(torch.autograd.Function):
+    """
+    Tries to handle half-precision derivatives in a randomized way that should
+    be more accurate for training than the default behavior.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor, dim: int):
+        ans = x.softmax(dim=dim)
+        # if x dtype is float16, x.softmax() returns a float32 because
+        # (presumably) that op does not support float16, and autocast
+        # is enabled.
+        if torch.is_autocast_enabled():
+            ans = ans.to(torch.float16)
+        ctx.save_for_backward(ans)
+        ctx.x_dtype = x.dtype
+        ctx.dim = dim
+        return ans
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        (ans,) = ctx.saved_tensors
+        with torch.cuda.amp.autocast(enabled=False):
+            ans_grad = ans_grad.to(torch.float32)
+            ans = ans.to(torch.float32)
+            x_grad = ans_grad * ans
+            x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True)
+            return x_grad, None
+
+
+def softmax(x: Tensor, dim: int):
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x.softmax(dim)
+
+    return SoftmaxFunction.apply(x, dim)
+
+
+class MaxEigLimiterFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        coeffs: Tensor,
+        direction: Tensor,
+        channel_dim: int,
+        grad_scale: float,
+    ) -> Tensor:
+        ctx.channel_dim = channel_dim
+        ctx.grad_scale = grad_scale
+        ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach())
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad, *args):
+        with torch.enable_grad():
+            (x_orig, coeffs, new_direction) = ctx.saved_tensors
+            x_orig.requires_grad = True
+            num_channels = x_orig.shape[ctx.channel_dim]
+            x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels)
+            new_direction.requires_grad = False
+            x = x - x.mean(dim=0)
+            x_var = (x ** 2).mean()
+            x_residual = x - coeffs * new_direction
+            x_residual_var = (x_residual ** 2).mean()
+            # `variance_proportion` is the proportion of the variance accounted for
+            # by the top eigen-direction.  This is to be minimized.
+            variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20)
+            variance_proportion.backward()
+        x_orig_grad = x_orig.grad
+        x_extra_grad = (
+            x_orig.grad
+            * ctx.grad_scale
+            * x_grad.norm()
+            / (x_orig_grad.norm() + 1.0e-20)
+        )
+        return x_grad + x_extra_grad.detach(), None, None, None, None
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+    eps_min: float
+    eps_max: float
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int = -1,  # CAUTION: see documentation.
+        eps: float = 0.25,
+        learn_eps: bool = True,
+        eps_min: float = -3.0,
+        eps_max: float = 3.0,
+    ) -> None:
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        if learn_eps:
+            self.eps = nn.Parameter(torch.tensor(eps).log().detach())
+        else:
+            self.register_buffer("eps", torch.tensor(eps).log().detach())
+        self.eps_min = eps_min
+        self.eps_max = eps_max
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+        eps = self.eps
+        if self.training and random.random() < 0.25:
+            # with probability 0.25, in training mode, clamp eps between the min
+            # and max; this will encourage it to learn parameters within the
+            # allowed range by making parameters that are outside the allowed
+            # range noisy.
+
+            # gradients to allow the parameter to get back into the allowed
+            # region if it happens to exit it.
+            eps = eps.clamp(min=self.eps_min, max=self.eps_max)
+        scales = (
+            torch.mean(x ** 2, dim=self.channel_dim, keepdim=True) + eps.exp()
+        ) ** -0.5
+        return x * scales
+
+
+def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear:
+    """
+    Behaves like a constructor of a modified version of nn.Linear
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Linear(*args, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(
+                ans.bias, -0.1 * initial_scale, 0.1 * initial_scale
+            )
+    return ans
+
+
+def ScaledConv1d(
+    *args,
+    initial_scale: float = 1.0,
+    kernel_size: int = 3,
+    padding: str = "same",
+    **kwargs,
+) -> nn.Conv1d:
+    """
+    Behaves like a constructor of a modified version of nn.Conv1d
+    that gives an easy way to set the default initial parameter scale.
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+    """
+    ans = nn.Conv1d(*args, kernel_size=kernel_size, padding=padding, **kwargs)
+    with torch.no_grad():
+        ans.weight[:] *= initial_scale
+        if ans.bias is not None:
+            torch.nn.init.uniform_(
+                ans.bias, -0.1 * initial_scale, 0.1 * initial_scale
+            )
+    return ans
+
+
+def TransposeScaledConv1d(
+    *args,
+    initial_scale: float = 1.0,
+    kernel_size: int = 3,
+    padding: str = "same",
+    **kwargs,
+) -> nn.Sequential:
+    """
+    Transpose -> ScaledConv1d
+    """
+    return nn.Sequential(
+        Transpose(),
+        ScaledConv1d(
+            *args,
+            initial_scale=initial_scale,
+            kernel_size=kernel_size,
+            padding=padding,
+            **kwargs,
+        ),
+    )
+
+
+def ScaledConv1dTranspose(
+    *args,
+    initial_scale: float = 1.0,
+    kernel_size: int = 3,
+    padding: str = "same",
+    **kwargs,
+) -> nn.Sequential:
+    """
+    Transpose -> ScaledConv1d
+    """
+    return nn.Sequential(
+        ScaledConv1d(
+            *args,
+            initial_scale=initial_scale,
+            kernel_size=kernel_size,
+            padding=padding,
+            **kwargs,
+        ),
+        Transpose(),
+    )
+
+
+def TransposeConv1d(
+    *args, kernel_size: int = 3, padding: str = "same", **kwargs
+) -> nn.Sequential:
+    """
+    Transpose -> Conv1d
+    """
+    return nn.Sequential(
+        Transpose(),
+        nn.Conv1d(*args, kernel_size=kernel_size, padding=padding, **kwargs),
+    )
+
+
+def Conv1dTranspose(
+    *args, kernel_size: int = 3, padding: str = "same", **kwargs
+) -> nn.Sequential:
+    """
+    ScaledConv1d -> Transpose
+    """
+    return nn.Sequential(
+        nn.Conv1d(*args, kernel_size=kernel_size, padding=padding, **kwargs),
+        Transpose(),
+    )
+
+
+class SRLinear(nn.Linear):
+    """https://arxiv.org/abs/2303.06296
+    Stabilizing Transformer Training by Preventing Attention Entropy Collapse
+    """
+
+    def __init__(self, in_features, out_features, bias=True, **kwargs):
+        super().__init__(in_features, out_features, bias=bias, **kwargs)
+        self.register_buffer(
+            "u", nn.functional.normalize(torch.randn(in_features), dim=0)
+        )
+        with torch.no_grad():
+            sigma = self.get_sigma()
+        self.register_buffer("spectral_norm", sigma)
+        self.sigma = nn.Parameter(torch.ones(1))
+
+    def get_sigma(self):
+        with torch.no_grad():
+            u = self.u
+            v = self.weight.mv(u)
+            v = nn.functional.normalize(v, dim=0)
+            u = self.weight.T.mv(v)
+            u = nn.functional.normalize(u, dim=0)
+            self.u.data.copy_(u)
+        return torch.einsum("c,cd,d->", v, self.weight, u)
+
+    def get_weight(self):
+        sigma = self.get_sigma()
+        if self.training:
+            self.spectral_norm.data.copy_(sigma)
+        weight = (self.sigma / sigma) * self.weight
+        return weight
+
+    def forward(self, x):
+        return nn.functional.linear(x, self.get_weight(), self.bias)
+
+
+class SRConv1d(SRLinear):
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        kernel_size,
+        stride: int = 1,
+        padding: str = "same",
+        bias: bool = True,
+        **kwargs,
+    ):
+        in_features = in_features * kernel_size
+        super().__init__(in_features, out_features, bias=bias, **kwargs)
+        nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.padding = padding
+
+    def forward(self, x):
+        in_features = self.in_features // self.kernel_size
+        weight = self.get_weight().view(
+            self.out_features, in_features, self.kernel_size
+        )
+        return nn.functional.conv1d(
+            x, weight, bias=self.bias, stride=self.stride, padding=self.padding
+        )
+
+
+def TransposeSRConv1d(
+    *args, kernel_size: int = 3, padding: str = "same", **kwargs
+) -> nn.Sequential:
+    """
+    Transpose -> SRConv1d
+    """
+    return nn.Sequential(
+        Transpose(),
+        SRConv1d(*args, kernel_size=kernel_size, padding=padding, **kwargs),
+    )
+
+
+def SRConv1dTranspose(
+    *args, kernel_size: int = 3, padding: str = "same", **kwargs
+) -> nn.Sequential:
+    """
+    SRConv1d -> Transpose
+    """
+    return nn.Sequential(
+        SRConv1d(*args, kernel_size=kernel_size, padding=padding, **kwargs),
+        Transpose(),
+    )
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           sign_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_positive and max_positive
+              are violated.
+           scale_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_abs and max_abs
+              are violated.
+           min_abs:  the minimum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+           max_abs:  the maximum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+          min_prob: determines the minimum probability with which we modify the
+             gradients for the {min,max}_positive and {min,max}_abs constraints,
+             on each forward().  This is done randomly to prevent all layers
+             from doing it at the same time.  Early in training we may use
+             higher probabilities than this; it will decay to this value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.04,
+        sign_gain_factor: float = 0.01,
+        scale_gain_factor: float = 0.02,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+        min_prob: float = 0.1,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        self.min_prob = min_prob
+        self.sign_gain_factor = sign_gain_factor
+        self.scale_gain_factor = scale_gain_factor
+
+        # count measures how many times the forward() function has been called.
+        # We occasionally sync this to a tensor called `count`, that exists to
+        # make sure it is synced to disk when we load and save the model.
+        self.cpu_count = 0
+        self.register_buffer("count", torch.tensor(0, dtype=torch.int64))
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or not x.requires_grad
+            or torch.jit.is_tracing()
+        ):
+            return _no_op(x)
+
+        count = self.cpu_count
+        self.cpu_count += 1
+
+        if random.random() < 0.01:
+            # Occasionally sync self.cpu_count with self.count.
+            # count affects the decay of 'prob'.  don't do this on every iter,
+            # because syncing with the GPU is slow.
+            self.cpu_count = max(self.cpu_count, self.count.item())
+            self.count.fill_(self.cpu_count)
+
+        # the prob of doing some work exponentially decreases from 0.5 till it hits
+        # a floor at min_prob (==0.1, by default)
+        prob = max(self.min_prob, 0.5 ** (1 + (count / 4000.0)))
+
+        if random.random() < prob:
+            sign_gain_factor = 0.5
+            if self.min_positive != 0.0 or self.max_positive != 1.0:
+                sign_factor = _compute_sign_factor(
+                    x,
+                    self.channel_dim,
+                    self.min_positive,
+                    self.max_positive,
+                    gain_factor=self.sign_gain_factor / prob,
+                    max_factor=self.max_factor,
+                )
+            else:
+                sign_factor = None
+
+            scale_factor = _compute_scale_factor(
+                x.detach(),
+                self.channel_dim,
+                min_abs=self.min_abs,
+                max_abs=self.max_abs,
+                gain_factor=self.scale_gain_factor / prob,
+                max_factor=self.max_factor,
+            )
+            return ActivationBalancerFunction.apply(
+                x,
+                scale_factor,
+                sign_factor,
+                self.channel_dim,
+            )
+        else:
+            return _no_op(x)
+
+
+def penalize_abs_values_gt(x: Tensor, limit: float, penalty: float) -> Tensor:
+    """
+    Returns x unmodified, but in backprop will put a penalty for the excess of
+    the absolute values of elements of x over the limit "limit".  E.g. if
+    limit == 10.0, then if x has any values over 10 it will get a penalty.
+
+    Caution: the value of this penalty will be affected by grad scaling used
+    in automatic mixed precision training.  For this reasons we use this,
+    it shouldn't really matter, or may even be helpful; we just use this
+    to disallow really implausible values of scores to be given to softmax.
+    """
+    x_sign = x.sign()
+    over_limit = (x.abs() - limit) > 0
+    # The following is a memory efficient way to penalize the absolute values of
+    # x that's over the limit.  (The memory efficiency comes when you think
+    # about which items torch needs to cache for the autograd, and which ones it
+    # can throw away).  The numerical value of aux_loss as computed here will
+    # actually be larger than it should be, by limit * over_limit.sum(), but it
+    # has the same derivative as the real aux_loss which is penalty * (x.abs() -
+    # limit).relu().
+    aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x)
+    # note: we don't do sum() here on aux)_loss, but it's as if we had done
+    # sum() due to how with_loss() works.
+    x = with_loss(x, aux_loss)
+    # you must use x for something, or this will be ineffective.
+    return x
+
+
+def _diag(x: Tensor):  # like .diag(), but works for tensors with 3 dims.
+    if x.ndim == 2:
+        return x.diag()
+    else:
+        (batch, dim, dim) = x.shape
+        x = x.reshape(batch, dim * dim)
+        x = x[:, :: dim + 1]
+        assert x.shape == (batch, dim)
+        return x
+
+
+def _whitening_metric(x: Tensor, num_groups: int):
+    """
+    Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of
+    of the centered feature covariance are the same within each group's covariance matrix
+    and also between groups.
+    Args:
+        x: a Tensor of shape (*, num_channels)
+     num_groups:  the number of groups of channels, a number >=1 that divides num_channels
+    Returns:
+        Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and
+    greater than 1.0 otherwise.
+    """
+    assert x.dtype != torch.float16
+    x = x.reshape(-1, x.shape[-1])
+    (num_frames, num_channels) = x.shape
+    assert num_channels % num_groups == 0
+    channels_per_group = num_channels // num_groups
+    x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1)
+    # x now has shape (num_groups, num_frames, channels_per_group)
+    # subtract the mean so we use the centered, not uncentered, covariance.
+    # My experience has been that when we "mess with the gradients" like this,
+    # it's better not do anything that tries to move the mean around, because
+    # that can easily cause instability.
+    x = x - x.mean(dim=1, keepdim=True)
+    # x_covar: (num_groups, channels_per_group, channels_per_group)
+    x_covar = torch.matmul(x.transpose(1, 2), x)
+    x_covar_mean_diag = _diag(x_covar).mean()
+    # the following expression is what we'd get if we took the matrix product
+    # of each covariance and measured the mean of its trace, i.e.
+    # the same as _diag(torch.matmul(x_covar, x_covar)).mean().
+    x_covarsq_mean_diag = (x_covar ** 2).sum() / (
+        num_groups * channels_per_group
+    )
+    # this metric will be >= 1.0; the larger it is, the less 'white' the data was.
+    metric = x_covarsq_mean_diag / (x_covar_mean_diag ** 2 + 1.0e-20)
+    return metric
+
+
+class WhiteningPenaltyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        num_groups: int,
+        whitening_limit: float,
+        grad_scale: float,
+    ) -> Tensor:
+        ctx.save_for_backward(x)
+        ctx.num_groups = num_groups
+        ctx.whitening_limit = whitening_limit
+        ctx.grad_scale = grad_scale
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor):
+        (x_orig,) = ctx.saved_tensors
+        with torch.enable_grad():
+            with torch.cuda.amp.autocast(enabled=False):
+                x_detached = x_orig.to(torch.float32).detach()
+                x_detached.requires_grad = True
+
+                metric = _whitening_metric(x_detached, ctx.num_groups)
+
+                if random.random() < 0.005 or __name__ == "__main__":
+                    logging.info(
+                        f"Whitening: num_groups={ctx.num_groups}, num_channels={x_orig.shape[-1]}, "
+                        f"metric={metric.item():.2f} vs. limit={ctx.whitening_limit}"
+                    )
+
+                (metric - ctx.whitening_limit).relu().backward()
+                penalty_grad = x_detached.grad
+                scale = ctx.grad_scale * (
+                    x_grad.to(torch.float32).norm()
+                    / (penalty_grad.norm() + 1.0e-20)
+                )
+                penalty_grad = penalty_grad * scale
+        return x_grad + penalty_grad.to(x_grad.dtype), None, None, None
+
+
+class Whiten(nn.Module):
+    def __init__(
+        self,
+        num_groups: int,
+        whitening_limit: float,
+        prob: Union[float, Tuple[float, float]],
+        grad_scale: float,
+    ):
+        """
+        Args:
+          num_groups: the number of groups to divide the channel dim into before
+            whitening.  We will attempt to make the feature covariance
+            within each group, after mean subtraction, as "white" as possible,
+            while having the same trace across all groups.
+         whitening_limit: a value greater than 1.0, that dictates how much
+           freedom we have to violate the constraints.  1.0 would mean perfectly
+           white, with exactly the same trace across groups; larger values
+           give more freedom.  E.g. 2.0.
+         prob: the probability with which we apply the gradient modification
+           (also affects the grad scale).  May be supplied as a float,
+           or as a pair (min_prob, max_prob)
+
+          grad_scale: determines the scale on the gradient term from this object,
+            relative to the rest of the gradient on the attention weights.
+            E.g. 0.02 (you may want to use smaller values than this if prob is large)
+        """
+        super(Whiten, self).__init__()
+        assert num_groups >= 1
+        assert whitening_limit >= 1
+        assert grad_scale >= 0
+        self.num_groups = num_groups
+        self.whitening_limit = whitening_limit
+        if isinstance(prob, float):
+            assert 0 < prob <= 1
+            self.prob = prob
+        else:
+            (self.min_prob, self.max_prob) = prob
+            assert 0 < self.min_prob < self.max_prob <= 1
+            self.prob = self.max_prob
+
+        self.grad_scale = grad_scale
+
+    def forward(self, x: Tensor) -> Tensor:
+        """
+        In the forward pass, this function just returns the input unmodified.
+        In the backward pass, it will modify the gradients to ensure that the
+        distribution in each group has close to (lambda times I) as the covariance
+        after mean subtraction, with the same lambda across groups.
+        For whitening_limit > 1, there will be more freedom to violate this
+        constraint.
+
+        Args:
+           x: the input of shape (*, num_channels)
+
+        Returns:
+            x, unmodified.   You should make sure
+        you use the returned value, or the graph will be freed
+        and nothing will happen in backprop.
+        """
+        if (
+            not x.requires_grad
+            or random.random() > self.prob
+            or self.grad_scale == 0
+        ):
+            return _no_op(x)
+        else:
+            if hasattr(self, "min_prob") and random.random() < 0.25:
+                # occasionally switch between min_prob and max_prob, based on whether
+                # we are above or below the threshold.
+                if (
+                    _whitening_metric(x.to(torch.float32), self.num_groups)
+                    > self.whitening_limit
+                ):
+                    # there would be a change to the grad.
+                    self.prob = self.max_prob
+                else:
+                    self.prob = self.min_prob
+
+            return WhiteningPenaltyFunction.apply(
+                x, self.num_groups, self.whitening_limit, self.grad_scale
+            )
+
+
+class WithLoss(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor, y: Tensor):
+        ctx.y_shape = y.shape
+        return x
+
+    @staticmethod
+    def backward(ctx, ans_grad: Tensor):
+        return ans_grad, torch.ones(
+            ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device
+        )
+
+
+def with_loss(x, y):
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x
+    # returns x but adds y.sum() to the loss function.
+    return WithLoss.apply(x, y)
+
+
+def _no_op(x: Tensor) -> Tensor:
+    if torch.jit.is_scripting() or torch.jit.is_tracing():
+        return x
+    else:
+        # a no-op function that will have a node in the autograd graph,
+        # to avoid certain bugs relating to backward hooks
+        return x.chunk(1, dim=-1)[0]
+
+
+class Identity(torch.nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, x):
+        return _no_op(x)
+
+
+class MaxEig(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to discourage
+    that any given direction in activation space accounts for more than
+    a specified proportion of the covariance (e.g. 0.2).
+
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           max_var_per_eig:  the maximum proportion of the variance of the
+               features/channels, after mean subtraction, that can come from
+               any given eigenvalue.
+           min_prob: the minimum probability with which we apply this during any invocation
+               of forward(), assuming last time we applied the constraint it was
+               not active; supplied for speed.
+           scale: determines the scale with which we modify the gradients, relative
+               to the existing / unmodified gradients
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        max_var_per_eig: float = 0.2,
+        min_prob: float = 0.01,
+        scale: float = 0.01,
+    ):
+        super(MaxEig, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.scale = scale
+        assert max_var_per_eig == 0.0 or max_var_per_eig > 1.0 / num_channels
+        self.max_var_per_eig = max_var_per_eig
+
+        # we figure out the dominant direction using the power method: starting with
+        # a random vector, keep multiplying by the covariance and renormalizing.
+        with torch.no_grad():
+            # arbitrary.. would use randn() but want to leave the rest of the model's
+            # random parameters unchanged for comparison
+            direction = torch.arange(num_channels).to(torch.float)
+            direction = direction / direction.norm()
+            self.register_buffer("max_eig_direction", direction)
+
+        self.min_prob = min_prob
+        # cur_prob is the current probability we'll use to apply the ActivationBalancer.
+        # We'll regress this towards prob, each time we try to apply it and it is not
+        # active.
+        self.cur_prob = 1.0
+
+    def forward(self, x: Tensor) -> Tensor:
+        if (
+            torch.jit.is_scripting()
+            or self.max_var_per_eig <= 0
+            or random.random() > self.cur_prob
+            or torch.jit.is_tracing()
+        ):
+            return _no_op(x)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            eps = 1.0e-20
+            orig_x = x
+            x = x.to(torch.float32)
+            with torch.no_grad():
+                x = x.transpose(self.channel_dim, -1).reshape(
+                    -1, self.num_channels
+                )
+                x = x - x.mean(dim=0)
+                new_direction, coeffs = self._find_direction_coeffs(
+                    x, self.max_eig_direction
+                )
+                x_var = (x ** 2).mean()
+                x_residual = x - coeffs * new_direction
+                x_residual_var = (x_residual ** 2).mean()
+
+                # `variance_proportion` is the proportion of the variance accounted for
+                # by the top eigen-direction.
+                variance_proportion = (x_var - x_residual_var) / (
+                    x_var + 1.0e-20
+                )
+
+                # ensure new direction is nonzero even if x == 0, by including `direction`.
+                self._set_direction(
+                    0.1 * self.max_eig_direction + new_direction
+                )
+
+            if random.random() < 0.01 or __name__ == "__main__":
+                logging.info(
+                    f"variance_proportion = {variance_proportion.item()}, shape={tuple(orig_x.shape)}, cur_prob={self.cur_prob}"
+                )
+
+            if variance_proportion >= self.max_var_per_eig:
+                # The constraint is active.  Note, we should quite rarely
+                # reach here, only near the beginning of training if we are
+                # starting to diverge, should this constraint be active.
+                cur_prob = self.cur_prob
+                self.cur_prob = (
+                    1.0  # next time, do the update with probability 1.0.
+                )
+                return MaxEigLimiterFunction.apply(
+                    orig_x, coeffs, new_direction, self.channel_dim, self.scale
+                )
+            else:
+                # let self.cur_prob exponentially approach self.min_prob, as
+                # long as the constraint is inactive.
+                self.cur_prob = 0.75 * self.cur_prob + 0.25 * self.min_prob
+                return orig_x
+
+    def _set_direction(self, direction: Tensor):
+        """
+        Sets self.max_eig_direction to a normalized version of `direction`
+        """
+        direction = direction.detach()
+        direction = direction / direction.norm()
+        direction_sum = direction.sum().item()
+        if direction_sum - direction_sum == 0:  # no inf/nan
+            self.max_eig_direction[:] = direction
+        else:
+            logging.info(
+                f"Warning: sum of direction in MaxEig is {direction_sum}, "
+                "num_channels={self.num_channels}, channel_dim={self.channel_dim}"
+            )
+
+    def _find_direction_coeffs(
+        self, x: Tensor, prev_direction: Tensor
+    ) -> Tuple[Tensor, Tensor, Tensor]:
+        """
+            Figure out (an approximation to) the proportion of the variance of a set of
+            feature vectors that can be attributed to the top eigen-direction.
+            Args:
+             x: a Tensor of shape (num_frames, num_channels), with num_frames > 1.
+          prev_direction:  a Tensor of shape (num_channels,), that is our previous estimate
+                   of the top eigen-direction, or a random direction if this is the first
+                   iteration.  Does not have to be normalized, but should be nonzero.
+
+        Returns: (cur_direction, coeffs), where:
+             cur_direction: a Tensor of shape (num_channels,) that is the current
+                estimate of the top eigen-direction.
+             coeffs: a Tensor of shape (num_frames, 1) that minimizes, or
+                approximately minimizes, (x - coeffs * cur_direction).norm()
+        """
+        (num_frames, num_channels) = x.shape
+        assert num_channels > 1 and num_frames > 1
+        assert prev_direction.shape == (num_channels,)
+        # `coeffs` are the coefficients of `prev_direction` in x.
+        # actually represent the coeffs up to a constant positive factor.
+        coeffs = (x * prev_direction).sum(dim=1, keepdim=True) + 1.0e-10
+        cur_direction = (x * coeffs).sum(dim=0) / (
+            (coeffs ** 2).sum() + 1.0e-20
+        )
+        return cur_direction, coeffs
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+
+        if requires_grad:
+            deriv = y * (1 - s) + s
+            # notes on derivative of x * sigmoid(x - 1):
+            # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
+            # min \simeq -0.043638.  Take floor as -0.043637 so it's a lower bund
+            # max \simeq 1.1990.   Take ceil to be 1.2 so it's an upper bound.
+            # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which
+            # floors), should be expectation-preserving.
+            floor = -0.043637
+            ceil = 1.2
+            d_scaled = (deriv - floor) * (
+                255.0 / (ceil - floor)
+            ) + torch.rand_like(deriv)
+            if __name__ == "__main__":
+                # for self-testing only.
+                assert d_scaled.min() >= 0.0
+                assert d_scaled.max() < 256.0
+            d_int = d_scaled.to(torch.uint8)
+            ctx.save_for_backward(d_int)
+        if x.dtype == torch.float16 or torch.is_autocast_enabled():
+            y = y.to(torch.float16)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.043637
+        ceil = 1.2
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            return x * torch.sigmoid(x - 1.0)
+        return DoubleSwishFunction.apply(x)
+
+
+def BalancedDoubleSwish(
+    d_model, channel_dim=-1, max_abs=10.0, min_prob=0.25
+) -> nn.Sequential:
+    """
+    ActivationBalancer -> DoubleSwish
+    """
+    balancer = ActivationBalancer(
+        d_model, channel_dim=channel_dim, max_abs=max_abs, min_prob=min_prob
+    )
+    return nn.Sequential(
+        balancer,
+        DoubleSwish(),
+    )
+
+
+def _test_max_eig():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = MaxEig(
+            num_channels, 1, 0.5, scale=0.1  # channel_dim  # max_var_per_eig
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad, atol=1.0e-02)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_whiten():
+    for proportion in [0.1, 0.5, 10.0]:
+        logging.info(f"_test_whiten(): proportion = {proportion}")
+        x = torch.randn(100, 128)
+        direction = torch.randn(128)
+        coeffs = torch.randn(100, 1)
+        x += proportion * direction * coeffs
+
+        x.requires_grad = True
+
+        num_channels = 128
+        m = Whiten(
+            1, 5.0, prob=1.0, grad_scale=0.1  # num_groups  # whitening_limit,
+        )  # grad_scale
+
+        for _ in range(4):
+            y = m(x)
+
+        y_grad = torch.randn_like(x)
+        y.backward(gradient=y_grad)
+
+        if proportion < 0.2:
+            assert torch.allclose(x.grad, y_grad)
+        elif proportion > 1.0:
+            assert not torch.allclose(x.grad, y_grad)
+
+
+def _test_activation_balancer_sign():
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * (
+        (2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0
+    )
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        probs.numel(),
+        channel_dim=0,
+        min_positive=0.05,
+        max_positive=0.95,
+        max_factor=0.2,
+        min_abs=0.0,
+    )
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+
+def _test_activation_balancer_magnitude():
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(
+        -1
+    )
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(
+        magnitudes.numel(),
+        channel_dim=0,
+        min_positive=0.0,
+        max_positive=1.0,
+        max_factor=0.2,
+        min_abs=0.2,
+        max_abs=0.8,
+        min_prob=1.0,
+    )
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x ** 2).mean().sqrt()
+    y_rms = (y ** 2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+def _test_double_swish_deriv():
+    x = torch.randn(10, 12, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    m = DoubleSwish()
+
+    tol = (1.2 - (-0.043637)) / 255.0
+    torch.autograd.gradcheck(m, x, atol=tol)
+
+    # for self-test.
+    x = torch.randn(1000, 1000, dtype=torch.double) * 3.0
+    x.requires_grad = True
+    y = m(x)
+
+
+def _test_softmax():
+    a = torch.randn(2, 10, dtype=torch.float64)
+    b = a.clone()
+    a.requires_grad = True
+    b.requires_grad = True
+    a.softmax(dim=1)[:, 0].sum().backward()
+    print("a grad = ", a.grad)
+    softmax(b, dim=1)[:, 0].sum().backward()
+    print("b grad = ", b.grad)
+    assert torch.allclose(a.grad, b.grad)
+
+
+if __name__ == "__main__":
+    logging.getLogger().setLevel(logging.INFO)
+    torch.set_num_threads(1)
+    torch.set_num_interop_threads(1)
+    _test_softmax()
+    _test_whiten()
+    _test_max_eig()
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()
+    _test_double_swish_deriv()

modules/scheduler.py

@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+# Copyright    2023                           (authors: Feiteng Li)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+
+from modules.optim import Eden
+
+
+def calc_lr(step, dim_embed, warmup_steps):
+    return dim_embed ** (-0.5) * min(
+        step ** (-0.5), step * warmup_steps ** (-1.5)
+    )
+
+
+class NoamScheduler(torch.optim.lr_scheduler._LRScheduler):
+    def __init__(
+        self,
+        base_lr: float,
+        optimizer: torch.optim.Optimizer,
+        dim_embed: int,
+        warmup_steps: int,
+        last_epoch: int = -1,
+        verbose: bool = False,
+    ) -> None:
+
+        self.dim_embed = dim_embed
+        self.base_lr = base_lr
+        self.warmup_steps = warmup_steps
+        self.num_param_groups = len(optimizer.param_groups)
+
+        super().__init__(optimizer, last_epoch, verbose)
+
+    def get_lr(self) -> float:
+        lr = self.base_lr * calc_lr(
+            self._step_count, self.dim_embed, self.warmup_steps
+        )
+        return [lr] * self.num_param_groups
+
+    def set_step(self, step: int):
+        self._step_count = step
+
+
+def get_scheduler(params, optimizer):
+    if params.scheduler_name.lower() == "eden":
+        scheduler = Eden(optimizer, 5000, 4, warmup_batches=params.warmup_steps)
+    elif params.scheduler_name.lower() == "noam":
+        scheduler = NoamScheduler(
+            params.base_lr,
+            optimizer,
+            params.decoder_dim,
+            warmup_steps=params.warmup_steps,
+        )
+        # scheduler.set_step(params.start_batch or params.batch_idx_train)
+    elif params.scheduler_name.lower() == "cosine":
+        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
+            params.warmup_steps,
+            optimizer,
+            eta_min=params.base_lr,
+        )
+    else:
+        raise NotImplementedError(f"{params.scheduler_name}")
+
+    return scheduler

modules/transformer.py

@@ -0,0 +1,683 @@
+import copy
+import numbers
+from functools import partial
+from typing import Any, Callable, List, Optional, Tuple, Union
+
+import torch
+from torch import Tensor, nn
+from torch.nn import functional as F
+
+from .activation import MultiheadAttention
+from .scaling import ActivationBalancer, BalancedDoubleSwish
+from .scaling import BasicNorm as _BasicNorm
+
+_shape_t = Union[int, List[int], torch.Size]
+
+
+class LayerNorm(nn.Module):
+    __constants__ = ["normalized_shape", "eps", "elementwise_affine"]
+    normalized_shape: Tuple[int, ...]
+    eps: float
+    elementwise_affine: bool
+
+    def __init__(
+        self,
+        normalized_shape: _shape_t,
+        eps: float = 1e-5,
+        elementwise_affine: bool = True,
+        device=None,
+        dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(LayerNorm, self).__init__()
+        if isinstance(normalized_shape, numbers.Integral):
+            # mypy error: incompatible types in assignment
+            normalized_shape = (normalized_shape,)  # type: ignore[assignment]
+        self.normalized_shape = tuple(normalized_shape)  # type: ignore[arg-type]
+        self.eps = eps
+        self.elementwise_affine = elementwise_affine
+        if self.elementwise_affine:
+            self.weight = nn.Parameter(
+                torch.empty(self.normalized_shape, **factory_kwargs)
+            )
+            self.bias = nn.Parameter(
+                torch.empty(self.normalized_shape, **factory_kwargs)
+            )
+        else:
+            self.register_parameter("weight", None)
+            self.register_parameter("bias", None)
+
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        if self.elementwise_affine:
+            nn.init.ones_(self.weight)
+            nn.init.zeros_(self.bias)
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            return (
+                F.layer_norm(
+                    input,
+                    self.normalized_shape,
+                    self.weight,
+                    self.bias,
+                    self.eps,
+                ),
+                embedding,
+            )
+
+        assert embedding is None
+        return F.layer_norm(
+            input, self.normalized_shape, self.weight, self.bias, self.eps
+        )
+
+    def extra_repr(self) -> str:
+        return (
+            "{normalized_shape}, eps={eps}, "
+            "elementwise_affine={elementwise_affine}".format(**self.__dict__)
+        )
+
+
+class AdaptiveLayerNorm(nn.Module):
+    r"""Adaptive Layer Normalization"""
+
+    def __init__(self, d_model, norm) -> None:
+        super(AdaptiveLayerNorm, self).__init__()
+        self.project_layer = nn.Linear(d_model, 2 * d_model)
+        self.norm = norm
+        self.d_model = d_model
+        self.eps = self.norm.eps
+
+    def forward(self, input: Tensor, embedding: Tensor = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            weight, bias = torch.split(
+                self.project_layer(embedding),
+                split_size_or_sections=self.d_model,
+                dim=-1,
+            )
+            return (weight * self.norm(input) + bias, embedding)
+
+        weight, bias = torch.split(
+            self.project_layer(embedding),
+            split_size_or_sections=self.d_model,
+            dim=-1,
+        )
+        return weight * self.norm(input) + bias
+
+
+class BasicNorm(_BasicNorm):
+    def __init__(
+        self,
+        d_model: int,
+        eps: float = 1e-5,
+        device=None,
+        dtype=None,
+    ):
+        super(BasicNorm, self).__init__(d_model, eps=eps)
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            return (
+                super(BasicNorm, self).forward(input),
+                embedding,
+            )
+
+        assert embedding is None
+        return super(BasicNorm, self).forward(input)
+
+
+class BalancedBasicNorm(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        eps: float = 1e-5,
+        device=None,
+        dtype=None,
+    ):
+        super(BalancedBasicNorm, self).__init__()
+        self.balancer = ActivationBalancer(
+            d_model,
+            channel_dim=-1,
+            min_positive=0.45,
+            max_positive=0.55,
+            max_abs=6.0,
+        )
+        self.norm = BasicNorm(d_model, eps, device=device, dtype=dtype)
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            return self.norm((self.balancer(input), embedding))
+
+        assert embedding is None
+        return self.norm(self.balancer(input))
+
+
+class IdentityNorm(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        eps: float = 1e-5,
+        device=None,
+        dtype=None,
+    ) -> None:
+        super(IdentityNorm, self).__init__()
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            return input
+
+        assert embedding is None
+        return input
+
+
+class TransformerEncoderLayer(nn.Module):
+    __constants__ = ["batch_first", "norm_first"]
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: Union[str, Callable[[Tensor], Tensor]] = F.relu,
+        batch_first: bool = False,
+        norm_first: bool = False,
+        device=None,
+        dtype=None,
+        linear1_self_attention_cls: nn.Module = nn.Linear,
+        linear2_self_attention_cls: nn.Module = nn.Linear,
+        linear1_feedforward_cls: nn.Module = nn.Linear,
+        linear2_feedforward_cls: nn.Module = nn.Linear,
+        layer_norm_cls: nn.Module = LayerNorm,
+        layer_norm_eps: float = 1e-5,
+        adaptive_layer_norm=False,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = MultiheadAttention(
+            d_model,
+            nhead,
+            dropout=dropout,
+            batch_first=batch_first,
+            linear1_cls=linear1_self_attention_cls,
+            linear2_cls=linear2_self_attention_cls,
+            **factory_kwargs,
+        )
+
+        # Implementation of Feedforward model
+        self.linear1 = linear1_feedforward_cls(
+            d_model, dim_feedforward, **factory_kwargs
+        )
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = linear2_feedforward_cls(
+            dim_feedforward, d_model, **factory_kwargs
+        )
+
+        self.norm_first = norm_first
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        # Legacy string support for activation function.
+        if isinstance(activation, str):
+            activation = _get_activation_fn(activation)
+        elif isinstance(activation, partial):
+            activation = activation(d_model)
+        elif activation == BalancedDoubleSwish:
+            activation = BalancedDoubleSwish(d_model)
+
+        # # We can't test self.activation in forward() in TorchScript,
+        # # so stash some information about it instead.
+        # if activation is F.relu or isinstance(activation, torch.nn.ReLU):
+        #     self.activation_relu_or_gelu = 1
+        # elif activation is F.gelu or isinstance(activation, torch.nn.GELU):
+        #     self.activation_relu_or_gelu = 2
+        # else:
+        #     self.activation_relu_or_gelu = 0
+        self.activation = activation
+
+        norm1 = layer_norm_cls(d_model, eps=layer_norm_eps, **factory_kwargs)
+        if layer_norm_cls == IdentityNorm:
+            norm2 = BalancedBasicNorm(
+                d_model, eps=layer_norm_eps, **factory_kwargs
+            )
+        else:
+            norm2 = layer_norm_cls(
+                d_model, eps=layer_norm_eps, **factory_kwargs
+            )
+
+        if adaptive_layer_norm:
+            self.norm1 = AdaptiveLayerNorm(d_model, norm1)
+            self.norm2 = AdaptiveLayerNorm(d_model, norm2)
+        else:
+            self.norm1 = norm1
+            self.norm2 = norm2
+
+    def __setstate__(self, state):
+        super(TransformerEncoderLayer, self).__setstate__(state)
+        if not hasattr(self, "activation"):
+            self.activation = F.relu
+
+    def forward(
+        self,
+        src: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            see the docs in Transformer class.
+        """
+        x, stage_embedding = src, None
+        is_src_tuple = False
+        if isinstance(src, tuple):
+            x, stage_embedding = src
+            is_src_tuple = True
+
+        if src_key_padding_mask is not None:
+            _skpm_dtype = src_key_padding_mask.dtype
+            if _skpm_dtype != torch.bool and not torch.is_floating_point(
+                src_key_padding_mask
+            ):
+                raise AssertionError(
+                    "only bool and floating types of key_padding_mask are supported"
+                )
+
+        if self.norm_first:
+            x = x + self._sa_block(
+                self.norm1(x, stage_embedding),
+                src_mask,
+                src_key_padding_mask,
+            )
+            x = x + self._ff_block(self.norm2(x, stage_embedding))
+        else:
+            x = self.norm1(
+                x + self._sa_block(x, src_mask, src_key_padding_mask),
+                stage_embedding,
+            )
+            x = self.norm2(x + self._ff_block(x), stage_embedding)
+
+        if is_src_tuple:
+            return (x, stage_embedding)
+        return x
+
+    def infer(
+        self,
+        src: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        past_kv: Optional[Tensor] = None,
+        use_cache: bool = False,
+    ):
+        x, stage_embedding = src, None
+        is_src_tuple = False
+        if isinstance(src, tuple):
+            x, stage_embedding = src
+            is_src_tuple = True
+
+        if src_key_padding_mask is not None:
+            _skpm_dtype = src_key_padding_mask.dtype
+            if _skpm_dtype != torch.bool and not torch.is_floating_point(
+                src_key_padding_mask
+            ):
+                raise AssertionError(
+                    "only bool and floating types of key_padding_mask are supported"
+                )
+
+        if self.norm_first:
+            x_attn_out, kv = self.self_attn.infer(
+                self.norm1(x, stage_embedding),
+                attn_mask=src_mask,
+                key_padding_mask=src_key_padding_mask,
+                need_weights=False,
+                past_kv=past_kv,
+                use_cache=use_cache,
+            )
+            x = x + x_attn_out
+            x = x + self._ff_block(self.norm2(x, stage_embedding))
+
+        if is_src_tuple:
+            return (x, stage_embedding)
+        return (x, kv)
+
+    # self-attention block
+    def _sa_block(
+        self,
+        x: Tensor,
+        attn_mask: Optional[Tensor],
+        key_padding_mask: Optional[Tensor],
+    ) -> Tensor:
+        x = self.self_attn(
+            x,
+            x,
+            x,
+            attn_mask=attn_mask,
+            key_padding_mask=key_padding_mask,
+            need_weights=False,
+        )[0]
+        return self.dropout1(x)
+
+    # feed forward block
+    def _ff_block(self, x: Tensor) -> Tensor:
+        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
+        return self.dropout2(x)
+
+
+class TransformerEncoder(nn.Module):
+    r"""TransformerEncoder is a stack of N encoder layers. Users can build the
+    BERT(https://arxiv.org/abs/1810.04805) model with corresponding parameters.
+
+    Args:
+        encoder_layer: an instance of the TransformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+        enable_nested_tensor: if True, input will automatically convert to nested tensor
+            (and convert back on output). This will improve the overall performance of
+            TransformerEncoder when padding rate is high. Default: ``True`` (enabled).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> transformer_encoder = TransformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = transformer_encoder(src)
+    """
+    __constants__ = ["norm"]
+
+    def __init__(self, encoder_layer, num_layers, norm=None):
+        super(TransformerEncoder, self).__init__()
+        self.layers = _get_clones(encoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = norm
+
+    def forward(
+        self,
+        src: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        return_layer_states: bool = False,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            return_layer_states: return layers' state (optional).
+
+        Shape:
+            see the docs in Transformer class.
+        """
+        if return_layer_states:
+            layer_states = []  # layers' output
+            output = src
+            for mod in self.layers:
+                output = mod(
+                    output,
+                    src_mask=mask,
+                    src_key_padding_mask=src_key_padding_mask,
+                )
+                layer_states.append(output[0])
+
+            if self.norm is not None:
+                output = self.norm(output)
+
+            return layer_states, output
+
+        output = src
+        for mod in self.layers:
+            output = mod(
+                output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+    def infer(
+        self,
+        src: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        return_layer_states: bool = False,
+        past_kv: Optional[Tensor] = None,
+        use_cache: bool = False,
+    ):
+        if past_kv is None:
+            past_length = 0
+            past_kv = tuple([None] * self.num_layers)
+        else:
+            past_length = past_kv[0][0].size(-2)
+        new_kv = () if use_cache else None
+        output = src
+        for mod, past_layer_kv in zip(self.layers, past_kv):
+            output, kv = mod.infer(
+                output, src_mask=mask, src_key_padding_mask=src_key_padding_mask, past_kv=past_layer_kv, use_cache=use_cache
+            )
+            if use_cache:
+                new_kv = new_kv + (kv,)
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output, new_kv
+
+
+class TransformerDecoderLayer(nn.Module):
+    __constants__ = ["batch_first", "norm_first"]
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: Union[str, Callable[[Tensor], Tensor]] = F.relu,
+        linear1_self_attention_cls: nn.Module = nn.Linear,
+        linear2_self_attention_cls: nn.Module = nn.Linear,
+        linear1_feedforward_cls: nn.Module = nn.Linear,
+        linear2_feedforward_cls: nn.Module = nn.Linear,
+        batch_first: bool = False,
+        norm_first: bool = False,
+        device=None,
+        dtype=None,
+        layer_norm_cls: nn.Module = LayerNorm,
+        layer_norm_eps: float = 1e-5,
+        adaptive_layer_norm=False,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(TransformerDecoderLayer, self).__init__()
+        self.self_attn = MultiheadAttention(
+            d_model,
+            nhead,
+            dropout=dropout,
+            batch_first=batch_first,
+            linear1_cls=linear1_self_attention_cls,
+            linear2_cls=linear2_self_attention_cls,
+            **factory_kwargs,
+        )
+        self.multihead_attn = MultiheadAttention(
+            d_model,
+            nhead,
+            dropout=dropout,
+            batch_first=batch_first,
+            linear1_cls=linear1_self_attention_cls,
+            linear2_cls=linear2_self_attention_cls,
+            **factory_kwargs,
+        )
+        # Implementation of Feedforward model
+        self.linear1 = linear1_feedforward_cls(
+            d_model, dim_feedforward, **factory_kwargs
+        )
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = linear2_feedforward_cls(
+            dim_feedforward, d_model, **factory_kwargs
+        )
+
+        self.norm_first = norm_first
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        # Legacy string support for activation function.
+        if isinstance(activation, str):
+            self.activation = _get_activation_fn(activation)
+        elif isinstance(activation, partial):
+            self.activation = activation(d_model)
+        elif activation == BalancedDoubleSwish:
+            self.activation = BalancedDoubleSwish(d_model)
+        else:
+            self.activation = activation
+
+        if adaptive_layer_norm:
+            norm1 = layer_norm_cls(
+                d_model, eps=layer_norm_eps, **factory_kwargs
+            )
+            norm2 = layer_norm_cls(
+                d_model, eps=layer_norm_eps, **factory_kwargs
+            )
+            norm3 = layer_norm_cls(
+                d_model, eps=layer_norm_eps, **factory_kwargs
+            )
+
+            self.norm1 = AdaptiveLayerNorm(d_model, norm1)
+            self.norm2 = AdaptiveLayerNorm(d_model, norm2)
+            self.norm3 = AdaptiveLayerNorm(d_model, norm3)
+        else:
+            self.norm1 = layer_norm_cls(
+                d_model, eps=layer_norm_eps, **factory_kwargs
+            )
+            self.norm2 = layer_norm_cls(
+                d_model, eps=layer_norm_eps, **factory_kwargs
+            )
+            if layer_norm_cls == IdentityNorm:
+                self.norm3 = BalancedBasicNorm(
+                    d_model, eps=layer_norm_eps, **factory_kwargs
+                )
+            else:
+                self.norm3 = layer_norm_cls(
+                    d_model, eps=layer_norm_eps, **factory_kwargs
+                )
+
+    def forward(
+        self,
+        tgt: Tensor,
+        memory: Tensor,
+        tgt_mask: Optional[Tensor] = None,
+        memory_mask: Optional[Tensor] = None,
+        tgt_key_padding_mask: Optional[Tensor] = None,
+        memory_key_padding_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        r"""Pass the inputs (and mask) through the decoder layer.
+
+        Args:
+            tgt: the sequence to the decoder layer (required).
+            memory: the sequence from the last layer of the encoder (required).
+            tgt_mask: the mask for the tgt sequence (optional).
+            memory_mask: the mask for the memory sequence (optional).
+            tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
+            memory_key_padding_mask: the mask for the memory keys per batch (optional).
+
+        Shape:
+            see the docs in Transformer class.
+        """
+        tgt_is_tuple = False
+        if isinstance(tgt, tuple):
+            x, stage_embedding = tgt
+            tgt_is_tuple = True
+        else:
+            x, stage_embedding = tgt, None
+
+        if self.norm_first:
+            x = x + self._sa_block(
+                self.norm1(x, stage_embedding), tgt_mask, tgt_key_padding_mask
+            )
+            x = x + self._mha_block(
+                self.norm2(x, stage_embedding),
+                memory,
+                memory_mask,
+                memory_key_padding_mask,
+            )
+            x = x + self._ff_block(self.norm3(x, stage_embedding))
+        else:
+            x = self.norm1(
+                x + self._sa_block(x, tgt_mask, tgt_key_padding_mask),
+                stage_embedding,
+            )
+            x = self.norm2(
+                x
+                + self._mha_block(
+                    x, memory, memory_mask, memory_key_padding_mask
+                ),
+                stage_embedding,
+            )
+            x = self.norm3(x + self._ff_block(x), stage_embedding)
+
+        if tgt_is_tuple:
+            return (x, stage_embedding)
+        return x
+
+    # self-attention block
+    def _sa_block(
+        self,
+        x: Tensor,
+        attn_mask: Optional[Tensor],
+        key_padding_mask: Optional[Tensor],
+    ) -> Tensor:
+        x = self.self_attn(
+            x,
+            x,
+            x,
+            attn_mask=attn_mask,
+            key_padding_mask=key_padding_mask,
+            need_weights=False,
+        )[0]
+        return self.dropout1(x)
+
+    # multihead attention block
+    def _mha_block(
+        self,
+        x: Tensor,
+        mem: Tensor,
+        attn_mask: Optional[Tensor],
+        key_padding_mask: Optional[Tensor],
+    ) -> Tensor:
+        x = self.multihead_attn(
+            x,
+            mem,
+            mem,
+            attn_mask=attn_mask,
+            key_padding_mask=key_padding_mask,
+            need_weights=False,
+        )[0]
+        return self.dropout2(x)
+
+    # feed forward block
+    def _ff_block(self, x: Tensor) -> Tensor:
+        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
+        return self.dropout3(x)
+
+
+def _get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+def _get_activation_fn(activation: str) -> Callable[[Tensor], Tensor]:
+    if activation == "relu":
+        return F.relu
+    elif activation == "gelu":
+        return F.gelu
+
+    raise RuntimeError(
+        "activation should be relu/gelu, not {}".format(activation)
+    )

nltk_data/tokenizers/punkt/PY3/README

@@ -0,0 +1,98 @@
+Pretrained Punkt Models -- Jan Strunk (New version trained after issues 313 and 514 had been corrected)
+
+Most models were prepared using the test corpora from Kiss and Strunk (2006). Additional models have
+been contributed by various people using NLTK for sentence boundary detection.
+
+For information about how to use these models, please confer the tokenization HOWTO:
+http://nltk.googlecode.com/svn/trunk/doc/howto/tokenize.html
+and chapter 3.8 of the NLTK book:
+http://nltk.googlecode.com/svn/trunk/doc/book/ch03.html#sec-segmentation
+
+There are pretrained tokenizers for the following languages:
+
+File                Language            Source                             Contents                Size of training corpus(in tokens)           Model contributed by
+=======================================================================================================================================================================
+czech.pickle        Czech               Multilingual Corpus 1 (ECI)        Lidove Noviny                   ~345,000                             Jan Strunk / Tibor Kiss
+                                                                           Literarni Noviny
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+danish.pickle       Danish              Avisdata CD-Rom Ver. 1.1. 1995     Berlingske Tidende              ~550,000                             Jan Strunk / Tibor Kiss
+                                        (Berlingske Avisdata, Copenhagen)  Weekend Avisen
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+dutch.pickle        Dutch               Multilingual Corpus 1 (ECI)        De Limburger                    ~340,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+english.pickle      English             Penn Treebank (LDC)                Wall Street Journal             ~469,000                             Jan Strunk / Tibor Kiss
+                    (American)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+estonian.pickle     Estonian            University of Tartu, Estonia       Eesti Ekspress                  ~359,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+finnish.pickle      Finnish             Finnish Parole Corpus, Finnish     Books and major national        ~364,000                             Jan Strunk / Tibor Kiss
+                                        Text Bank (Suomen Kielen           newspapers
+                                        Tekstipankki)
+                                        Finnish Center for IT Science
+                                        (CSC)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+french.pickle       French              Multilingual Corpus 1 (ECI)        Le Monde                        ~370,000                             Jan Strunk / Tibor Kiss
+                    (European)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+german.pickle       German              Neue Zürcher Zeitung AG            Neue Zürcher Zeitung            ~847,000                             Jan Strunk / Tibor Kiss
+                    (Switzerland)       CD-ROM
+                    (Uses "ss"
+                     instead of "ß")
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+greek.pickle        Greek               Efstathios Stamatatos              To Vima (TO BHMA)               ~227,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+italian.pickle      Italian             Multilingual Corpus 1 (ECI)        La Stampa, Il Mattino           ~312,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+norwegian.pickle    Norwegian           Centre for Humanities              Bergens Tidende                 ~479,000                             Jan Strunk / Tibor Kiss
+                    (Bokmål and         Information Technologies,
+                     Nynorsk)           Bergen
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+polish.pickle       Polish              Polish National Corpus             Literature, newspapers, etc.  ~1,000,000                             Krzysztof Langner
+                                        (http://www.nkjp.pl/)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+portuguese.pickle   Portuguese          CETENFolha Corpus                  Folha de São Paulo              ~321,000                             Jan Strunk / Tibor Kiss
+                    (Brazilian)         (Linguateca)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+slovene.pickle      Slovene             TRACTOR                            Delo                            ~354,000                             Jan Strunk / Tibor Kiss
+                                        Slovene Academy for Arts
+                                        and Sciences
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+spanish.pickle      Spanish             Multilingual Corpus 1 (ECI)        Sur                             ~353,000                             Jan Strunk / Tibor Kiss
+                    (European)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+swedish.pickle      Swedish             Multilingual Corpus 1 (ECI)        Dagens Nyheter                  ~339,000                             Jan Strunk / Tibor Kiss
+                                                                           (and some other texts)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+turkish.pickle      Turkish             METU Turkish Corpus                Milliyet                        ~333,000                             Jan Strunk / Tibor Kiss
+                                        (Türkçe Derlem Projesi)
+                                        University of Ankara
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+
+The corpora contained about 400,000 tokens on average and mostly consisted of newspaper text converted to
+Unicode using the codecs module.
+
+Kiss, Tibor and Strunk, Jan (2006): Unsupervised Multilingual Sentence Boundary Detection.
+Computational Linguistics 32: 485-525.
+
+---- Training Code ----
+
+# import punkt
+import nltk.tokenize.punkt
+
+# Make a new Tokenizer
+tokenizer = nltk.tokenize.punkt.PunktSentenceTokenizer()
+
+# Read in training corpus (one example: Slovene)
+import codecs
+text = codecs.open("slovene.plain","Ur","iso-8859-2").read()
+
+# Train tokenizer
+tokenizer.train(text)
+
+# Dump pickled tokenizer
+import pickle
+out = open("slovene.pickle","wb")
+pickle.dump(tokenizer, out)
+out.close()
+
+---------

nltk_data/tokenizers/punkt/PY3/czech.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64b0734b6fbe8e8d7cac79f48d1dd9f853824e57c4e3594dadd74ba2c1d97f50
+size 1119050

nltk_data/tokenizers/punkt/PY3/danish.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6189c7dd254e29e2bd406a7f6a4336297c8953214792466a790ea4444223ceb3
+size 1191710

nltk_data/tokenizers/punkt/PY3/dutch.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fda0d6a13f02e8898daec7fe923da88e25abe081bcfa755c0e015075c215fe4c
+size 693759

nltk_data/tokenizers/punkt/PY3/english.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cad3758596392364e3be9803dbd7ebeda384b68937b488a01365f5551bb942c
+size 406697

nltk_data/tokenizers/punkt/PY3/estonian.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b364f72538d17b146a98009ad239a8096ce6c0a8b02958c0bc776ecd0c58a25f
+size 1499502

nltk_data/tokenizers/punkt/PY3/finnish.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a4b5ff5500ee851c456f9dd40d5fc0d8c1859c88eb3178de1317d26b7d22833
+size 1852226

nltk_data/tokenizers/punkt/PY3/french.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28e3a4cd2971989b3cb9fd3433a6f15d17981e464db2be039364313b5de94f29
+size 553575

nltk_data/tokenizers/punkt/PY3/german.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddcbbe85e2042a019b1a6e37fd8c153286c38ba201fae0f5bfd9a3f74abae25c
+size 1463575

nltk_data/tokenizers/punkt/PY3/greek.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85dabc44ab90a5f208ef37ff6b4892ebe7e740f71fb4da47cfd95417ca3e22fd
+size 876006

nltk_data/tokenizers/punkt/PY3/italian.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68a94007b1e4ffdc4d1a190185ca5442c3dafeb17ab39d30329e84cd74a43947
+size 615089

nltk_data/tokenizers/punkt/PY3/malayalam.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f8cf58acbdb7f472ac40affc13663be42dafb47c15030c11ade0444c9e0e53d
+size 221207

nltk_data/tokenizers/punkt/PY3/norwegian.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ff7a46d1438b311457d15d7763060b8d3270852c1850fd788c5cee194dc4a1d
+size 1181271

nltk_data/tokenizers/punkt/PY3/polish.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:624900ae3ddfb4854a98c5d3b8b1c9bb719975f33fee61ce1441dab9f8a00718
+size 1738386

nltk_data/tokenizers/punkt/PY3/portuguese.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02a0b7b25c3c7471e1791b66a31bbb530afbb0160aee4fcecf0107652067b4a1
+size 611919

nltk_data/tokenizers/punkt/PY3/russian.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:549762f8190024d89b511472df21a3a135eee5d9233e63ac244db737c2c61d7e
+size 33020

nltk_data/tokenizers/punkt/PY3/slovene.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52ef2cc0ed27d79b3aa635cbbc40ad811883a75a4b8a8be1ae406972870fd864
+size 734444

nltk_data/tokenizers/punkt/PY3/spanish.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:164a50fadc5a49f8ec7426eae11d3111ee752b48a3ef373d47745011192a5984
+size 562337

nltk_data/tokenizers/punkt/PY3/swedish.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0f7d538bfd5266633b09e842cd92e9e0ac10f1d923bf211e1497972ddc47318
+size 979681

nltk_data/tokenizers/punkt/PY3/turkish.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae68ef5863728ac5332e87eb1f6bae772ff32a13a4caa2b01a5c68103e853c5b
+size 1017038

nltk_data/tokenizers/punkt/README

@@ -0,0 +1,98 @@
+Pretrained Punkt Models -- Jan Strunk (New version trained after issues 313 and 514 had been corrected)
+
+Most models were prepared using the test corpora from Kiss and Strunk (2006). Additional models have
+been contributed by various people using NLTK for sentence boundary detection.
+
+For information about how to use these models, please confer the tokenization HOWTO:
+http://nltk.googlecode.com/svn/trunk/doc/howto/tokenize.html
+and chapter 3.8 of the NLTK book:
+http://nltk.googlecode.com/svn/trunk/doc/book/ch03.html#sec-segmentation
+
+There are pretrained tokenizers for the following languages:
+
+File                Language            Source                             Contents                Size of training corpus(in tokens)           Model contributed by
+=======================================================================================================================================================================
+czech.pickle        Czech               Multilingual Corpus 1 (ECI)        Lidove Noviny                   ~345,000                             Jan Strunk / Tibor Kiss
+                                                                           Literarni Noviny
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+danish.pickle       Danish              Avisdata CD-Rom Ver. 1.1. 1995     Berlingske Tidende              ~550,000                             Jan Strunk / Tibor Kiss
+                                        (Berlingske Avisdata, Copenhagen)  Weekend Avisen
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+dutch.pickle        Dutch               Multilingual Corpus 1 (ECI)        De Limburger                    ~340,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+english.pickle      English             Penn Treebank (LDC)                Wall Street Journal             ~469,000                             Jan Strunk / Tibor Kiss
+                    (American)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+estonian.pickle     Estonian            University of Tartu, Estonia       Eesti Ekspress                  ~359,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+finnish.pickle      Finnish             Finnish Parole Corpus, Finnish     Books and major national        ~364,000                             Jan Strunk / Tibor Kiss
+                                        Text Bank (Suomen Kielen           newspapers
+                                        Tekstipankki)
+                                        Finnish Center for IT Science
+                                        (CSC)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+french.pickle       French              Multilingual Corpus 1 (ECI)        Le Monde                        ~370,000                             Jan Strunk / Tibor Kiss
+                    (European)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+german.pickle       German              Neue Zürcher Zeitung AG            Neue Zürcher Zeitung            ~847,000                             Jan Strunk / Tibor Kiss
+                    (Switzerland)       CD-ROM
+                    (Uses "ss"
+                     instead of "ß")
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+greek.pickle        Greek               Efstathios Stamatatos              To Vima (TO BHMA)               ~227,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+italian.pickle      Italian             Multilingual Corpus 1 (ECI)        La Stampa, Il Mattino           ~312,000                             Jan Strunk / Tibor Kiss
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+norwegian.pickle    Norwegian           Centre for Humanities              Bergens Tidende                 ~479,000                             Jan Strunk / Tibor Kiss
+                    (Bokmål and         Information Technologies,
+                     Nynorsk)           Bergen
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+polish.pickle       Polish              Polish National Corpus             Literature, newspapers, etc.  ~1,000,000                             Krzysztof Langner
+                                        (http://www.nkjp.pl/)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+portuguese.pickle   Portuguese          CETENFolha Corpus                  Folha de São Paulo              ~321,000                             Jan Strunk / Tibor Kiss
+                    (Brazilian)         (Linguateca)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+slovene.pickle      Slovene             TRACTOR                            Delo                            ~354,000                             Jan Strunk / Tibor Kiss
+                                        Slovene Academy for Arts
+                                        and Sciences
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+spanish.pickle      Spanish             Multilingual Corpus 1 (ECI)        Sur                             ~353,000                             Jan Strunk / Tibor Kiss
+                    (European)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+swedish.pickle      Swedish             Multilingual Corpus 1 (ECI)        Dagens Nyheter                  ~339,000                             Jan Strunk / Tibor Kiss
+                                                                           (and some other texts)
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+turkish.pickle      Turkish             METU Turkish Corpus                Milliyet                        ~333,000                             Jan Strunk / Tibor Kiss
+                                        (Türkçe Derlem Projesi)
+                                        University of Ankara
+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
+
+The corpora contained about 400,000 tokens on average and mostly consisted of newspaper text converted to
+Unicode using the codecs module.
+
+Kiss, Tibor and Strunk, Jan (2006): Unsupervised Multilingual Sentence Boundary Detection.
+Computational Linguistics 32: 485-525.
+
+---- Training Code ----
+
+# import punkt
+import nltk.tokenize.punkt
+
+# Make a new Tokenizer
+tokenizer = nltk.tokenize.punkt.PunktSentenceTokenizer()
+
+# Read in training corpus (one example: Slovene)
+import codecs
+text = codecs.open("slovene.plain","Ur","iso-8859-2").read()
+
+# Train tokenizer
+tokenizer.train(text)
+
+# Dump pickled tokenizer
+import pickle
+out = open("slovene.pickle","wb")
+pickle.dump(tokenizer, out)
+out.close()
+
+---------

nltk_data/tokenizers/punkt/czech.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c085f6283bed0f1390d36a55d126ccc29c9b4dfcd2705e862b1711b7c6bb5ab
+size 1424691

nltk_data/tokenizers/punkt/danish.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df8366ad67db22b1f838cd63fcc589a6006faf66d7a46be5312d9c487ce2c811
+size 1427491

nltk_data/tokenizers/punkt/dutch.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12f46024d3c840529b56ac2a3118b80b8dc77705734bcdd71ff7c46f5808395e
+size 839761