Create onnx_inference.py
Browse files- onnx_inference.py +103 -0
onnx_inference.py
ADDED
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import json
|
3 |
+
from huggingface_hub import hf_hub_download
|
4 |
+
import re
|
5 |
+
import emoji
|
6 |
+
from transformers import BertTokenizer
|
7 |
+
import onnxruntime as ort
|
8 |
+
|
9 |
+
def preprocess_text(text):
|
10 |
+
"""Preprocess the input text to match training conditions."""
|
11 |
+
text = re.sub(r'u/\w+', '[USER]', text)
|
12 |
+
text = re.sub(r'r/\w+', '[SUBREDDIT]', text)
|
13 |
+
text = re.sub(r'http[s]?://\S+', '[URL]', text)
|
14 |
+
text = emoji.demojize(text, delimiters=(" ", " "))
|
15 |
+
text = text.lower()
|
16 |
+
return text
|
17 |
+
|
18 |
+
def load_model_and_resources():
|
19 |
+
"""Load the ONNX model, tokenizer, emotion labels, and thresholds from Hugging Face."""
|
20 |
+
repo_id = "logasanjeev/goemotions-bert"
|
21 |
+
|
22 |
+
try:
|
23 |
+
tokenizer = BertTokenizer.from_pretrained(repo_id)
|
24 |
+
except Exception as e:
|
25 |
+
raise RuntimeError(f"Error loading tokenizer: {str(e)}")
|
26 |
+
|
27 |
+
try:
|
28 |
+
model_path = hf_hub_download(repo_id=repo_id, filename="model.onnx")
|
29 |
+
session = ort.InferenceSession(model_path)
|
30 |
+
except Exception as e:
|
31 |
+
raise RuntimeError(f"Error loading ONNX model: {str(e)}")
|
32 |
+
|
33 |
+
try:
|
34 |
+
thresholds_file = hf_hub_download(repo_id=repo_id, filename="optimized_thresholds.json")
|
35 |
+
with open(thresholds_file, "r") as f:
|
36 |
+
thresholds_data = json.load(f)
|
37 |
+
if not (isinstance(thresholds_data, dict) and "emotion_labels" in thresholds_data and "thresholds" in thresholds_data):
|
38 |
+
raise ValueError("Unexpected format in optimized_thresholds.json. Expected a dictionary with keys 'emotion_labels' and 'thresholds'.")
|
39 |
+
emotion_labels = thresholds_data["emotion_labels"]
|
40 |
+
thresholds = thresholds_data["thresholds"]
|
41 |
+
except Exception as e:
|
42 |
+
raise RuntimeError(f"Error loading thresholds: {str(e)}")
|
43 |
+
|
44 |
+
return session, tokenizer, emotion_labels, thresholds
|
45 |
+
|
46 |
+
SESSION, TOKENIZER, EMOTION_LABELS, THRESHOLDS = None, None, None, None
|
47 |
+
|
48 |
+
def predict_emotions(text):
|
49 |
+
"""Predict emotions for the given text using the GoEmotions BERT ONNX model.
|
50 |
+
|
51 |
+
Args:
|
52 |
+
text (str): The input text to analyze.
|
53 |
+
|
54 |
+
Returns:
|
55 |
+
tuple: (predictions, processed_text)
|
56 |
+
- predictions (str): Formatted string of predicted emotions and their confidence scores.
|
57 |
+
- processed_text (str): The preprocessed input text.
|
58 |
+
"""
|
59 |
+
global SESSION, TOKENIZER, EMOTION_LABELS, THRESHOLDS
|
60 |
+
|
61 |
+
if SESSION is None:
|
62 |
+
SESSION, TOKENIZER, EMOTION_LABELS, THRESHOLDS = load_model_and_resources()
|
63 |
+
|
64 |
+
processed_text = preprocess_text(text)
|
65 |
+
|
66 |
+
encodings = TOKENIZER(
|
67 |
+
processed_text,
|
68 |
+
padding='max_length',
|
69 |
+
truncation=True,
|
70 |
+
max_length=128,
|
71 |
+
return_tensors='np'
|
72 |
+
)
|
73 |
+
|
74 |
+
inputs = {
|
75 |
+
'input_ids': encodings['input_ids'].astype(np.int64),
|
76 |
+
'attention_mask': encodings['attention_mask'].astype(np.int64)
|
77 |
+
}
|
78 |
+
|
79 |
+
logits = SESSION.run(None, inputs)[0][0]
|
80 |
+
logits = 1 / (1 + np.exp(-logits)) # Sigmoid
|
81 |
+
|
82 |
+
predictions = []
|
83 |
+
for i, (logit, thresh) in enumerate(zip(logits, THRESHOLDS)):
|
84 |
+
if logit >= thresh:
|
85 |
+
predictions.append((EMOTION_LABELS[i], round(logit, 4)))
|
86 |
+
|
87 |
+
predictions.sort(key=lambda x: x[1], reverse=True)
|
88 |
+
|
89 |
+
result = "\n".join([f"{emotion}: {confidence:.4f}" for emotion, confidence in predictions]) or "No emotions predicted."
|
90 |
+
return result, processed_text
|
91 |
+
|
92 |
+
if __name__ == "__main__":
|
93 |
+
import argparse
|
94 |
+
|
95 |
+
parser = argparse.ArgumentParser(description="Predict emotions using the GoEmotions BERT ONNX model.")
|
96 |
+
parser.add_argument("text", type=str, help="The input text to analyze for emotions.")
|
97 |
+
args = parser.parse_args()
|
98 |
+
|
99 |
+
result, processed = predict_emotions(args.text)
|
100 |
+
print(f"Input: {args.text}")
|
101 |
+
print(f"Processed: {processed}")
|
102 |
+
print("Predicted Emotions:")
|
103 |
+
print(result)
|