Julien Simon
commited on
Commit
·
b536abf
1
Parent(s):
ead8a38
Training in progress, epoch 3
Browse files- pytorch_model.bin +1 -1
- train-xlm.py +114 -0
pytorch_model.bin
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 3114359925
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d62f1c5ab88bf2f7b3820b4b411f1b51a423796b4c6ad6fa37f8e21629d5c28d
|
| 3 |
size 3114359925
|
train-xlm.py
ADDED
|
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import evaluate
|
| 2 |
+
import numpy as np
|
| 3 |
+
from datasets import load_dataset
|
| 4 |
+
from transformers import (
|
| 5 |
+
AutoTokenizer,
|
| 6 |
+
AutoModelForSequenceClassification,
|
| 7 |
+
Trainer,
|
| 8 |
+
TrainingArguments,
|
| 9 |
+
)
|
| 10 |
+
|
| 11 |
+
dataset_id = "google/fleurs"
|
| 12 |
+
model_id = "facebook/xlm-v-base"
|
| 13 |
+
metric_name = "accuracy"
|
| 14 |
+
|
| 15 |
+
# Keep only the raw transcription and the language id (which we'll use as label)
|
| 16 |
+
columns_to_remove = [
|
| 17 |
+
"audio",
|
| 18 |
+
"id",
|
| 19 |
+
"num_samples",
|
| 20 |
+
"path",
|
| 21 |
+
"transcription",
|
| 22 |
+
"gender",
|
| 23 |
+
"language",
|
| 24 |
+
"lang_group_id",
|
| 25 |
+
]
|
| 26 |
+
|
| 27 |
+
train, val = load_dataset(
|
| 28 |
+
dataset_id, "all", split=["train", "validation"], ignore_verifications=True
|
| 29 |
+
)
|
| 30 |
+
|
| 31 |
+
# Build the label2id and id2label dictionaries
|
| 32 |
+
|
| 33 |
+
unique_langs = set()
|
| 34 |
+
label2id = {}
|
| 35 |
+
id2label = {}
|
| 36 |
+
for lang, lang_id in zip(val["language"], val["lang_id"]):
|
| 37 |
+
if lang not in unique_langs:
|
| 38 |
+
unique_langs.add(lang)
|
| 39 |
+
id2label[lang_id] = lang
|
| 40 |
+
label2id[lang] = lang_id
|
| 41 |
+
|
| 42 |
+
id2label = dict(sorted(id2label.items(), key=lambda item: item[0]))
|
| 43 |
+
label2id = dict(sorted(label2id.items(), key=lambda item: item[1]))
|
| 44 |
+
|
| 45 |
+
train = train.remove_columns(columns_to_remove)
|
| 46 |
+
val = val.remove_columns(columns_to_remove)
|
| 47 |
+
train = train.rename_column("raw_transcription", "text")
|
| 48 |
+
val = val.rename_column("raw_transcription", "text")
|
| 49 |
+
train = train.rename_column("lang_id", "label")
|
| 50 |
+
val = val.rename_column("lang_id", "label")
|
| 51 |
+
|
| 52 |
+
train = train.shuffle(seed=42)
|
| 53 |
+
val = val.shuffle(seed=42)
|
| 54 |
+
|
| 55 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
def preprocess(data):
|
| 59 |
+
return tokenizer(data["text"], truncation=True)
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
processed_train = train.map(preprocess, batched=True)
|
| 63 |
+
processed_val = val.map(preprocess, batched=True)
|
| 64 |
+
|
| 65 |
+
print(processed_train)
|
| 66 |
+
print(processed_val)
|
| 67 |
+
|
| 68 |
+
# Fine-tune the model
|
| 69 |
+
|
| 70 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
| 71 |
+
model_id,
|
| 72 |
+
num_labels=len(id2label),
|
| 73 |
+
label2id=label2id,
|
| 74 |
+
id2label=id2label,
|
| 75 |
+
ignore_mismatched_sizes=True,
|
| 76 |
+
)
|
| 77 |
+
|
| 78 |
+
args = TrainingArguments(
|
| 79 |
+
"xlm-v-base-language-id",
|
| 80 |
+
learning_rate=3e-5,
|
| 81 |
+
warmup_ratio=0.1,
|
| 82 |
+
per_device_train_batch_size=16,
|
| 83 |
+
gradient_accumulation_steps=4,
|
| 84 |
+
per_device_eval_batch_size=16,
|
| 85 |
+
num_train_epochs=5,
|
| 86 |
+
load_best_model_at_end=True,
|
| 87 |
+
metric_for_best_model=metric_name,
|
| 88 |
+
evaluation_strategy="epoch",
|
| 89 |
+
save_strategy="epoch",
|
| 90 |
+
logging_steps=10,
|
| 91 |
+
fp16=True,
|
| 92 |
+
push_to_hub=True,
|
| 93 |
+
)
|
| 94 |
+
|
| 95 |
+
metric = evaluate.load(metric_name)
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
def compute_metrics(eval_pred):
|
| 99 |
+
predictions = np.argmax(eval_pred.predictions, axis=1)
|
| 100 |
+
return metric.compute(predictions=predictions, references=eval_pred.label_ids)
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
trainer = Trainer(
|
| 104 |
+
model,
|
| 105 |
+
args,
|
| 106 |
+
train_dataset=processed_train,
|
| 107 |
+
eval_dataset=processed_val,
|
| 108 |
+
tokenizer=tokenizer,
|
| 109 |
+
compute_metrics=compute_metrics,
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
trainer.train()
|
| 113 |
+
|
| 114 |
+
trainer.save_model("./my_model")
|