Spaces:
Sleeping
Sleeping
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import FastAPI, HTTPException
|
2 |
+
from pydantic import BaseModel
|
3 |
+
from langdetect import detect
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
from transformers import (
|
7 |
+
DistilBertTokenizer, DistilBertModel,
|
8 |
+
AutoTokenizer, AutoModel
|
9 |
+
)
|
10 |
+
|
11 |
+
# ==== Model Classes ====
|
12 |
+
|
13 |
+
class ToxicBERT(nn.Module):
|
14 |
+
def __init__(self):
|
15 |
+
super().__init__()
|
16 |
+
self.bert = DistilBertModel.from_pretrained("distilbert-base-uncased")
|
17 |
+
self.dropout = nn.Dropout(0.3)
|
18 |
+
self.classifier = nn.Linear(self.bert.config.hidden_size, 6)
|
19 |
+
|
20 |
+
def forward(self, input_ids, attention_mask):
|
21 |
+
output = self.bert(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state[:, 0]
|
22 |
+
return self.classifier(self.dropout(output))
|
23 |
+
|
24 |
+
class HinglishToxicClassifier(nn.Module):
|
25 |
+
def __init__(self):
|
26 |
+
super().__init__()
|
27 |
+
self.bert = AutoModel.from_pretrained("xlm-roberta-base")
|
28 |
+
hidden_size = self.bert.config.hidden_size
|
29 |
+
|
30 |
+
self.pool = lambda hidden: torch.cat([
|
31 |
+
hidden.mean(dim=1),
|
32 |
+
hidden.max(dim=1).values
|
33 |
+
], dim=1)
|
34 |
+
|
35 |
+
self.bottleneck = nn.Sequential(
|
36 |
+
nn.Linear(2 * hidden_size, hidden_size),
|
37 |
+
nn.ReLU(),
|
38 |
+
nn.Dropout(0.2)
|
39 |
+
)
|
40 |
+
self.classifier = nn.Linear(hidden_size, 2)
|
41 |
+
|
42 |
+
def forward(self, input_ids, attention_mask):
|
43 |
+
hidden = self.bert(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state
|
44 |
+
pooled = self.pool(hidden)
|
45 |
+
x = self.bottleneck(pooled)
|
46 |
+
return self.classifier(x)
|
47 |
+
|
48 |
+
# ==== Load Tokenizers ====
|
49 |
+
english_tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
|
50 |
+
hinglish_tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")
|
51 |
+
|
52 |
+
# ==== Load Models from Hugging Face Hub ====
|
53 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
54 |
+
|
55 |
+
english_model = ToxicBERT()
|
56 |
+
eng_url = "https://huggingface.co/koyu008/English_Toxic_Classifier/resolve/main/bert_toxic_classifier.pt"
|
57 |
+
english_model.load_state_dict(torch.hub.load_state_dict_from_url(eng_url, map_location=device))
|
58 |
+
english_model.eval().to(device)
|
59 |
+
|
60 |
+
hinglish_model = HinglishToxicClassifier()
|
61 |
+
hin_url = "https://huggingface.co/koyu008/HInglish_comment_classifier/resolve/main/best_hinglish_model.pt"
|
62 |
+
hinglish_model.load_state_dict(torch.hub.load_state_dict_from_url(hin_url, map_location=device))
|
63 |
+
hinglish_model.eval().to(device)
|
64 |
+
|
65 |
+
# ==== FastAPI setup ====
|
66 |
+
|
67 |
+
app = FastAPI()
|
68 |
+
|
69 |
+
class InputText(BaseModel):
|
70 |
+
text: str
|
71 |
+
|
72 |
+
@app.post("/predict")
|
73 |
+
def predict(input: InputText):
|
74 |
+
text = input.text.strip()
|
75 |
+
if not text:
|
76 |
+
raise HTTPException(status_code=400, detail="Input text cannot be empty")
|
77 |
+
|
78 |
+
# Language detection
|
79 |
+
try:
|
80 |
+
lang = detect(text)
|
81 |
+
except:
|
82 |
+
lang = "und"
|
83 |
+
|
84 |
+
if lang == "en":
|
85 |
+
model = english_model
|
86 |
+
tokenizer = english_tokenizer
|
87 |
+
labels = ["toxic", "severe toxic", "obscene", "threat", "insult", "identity hate"]
|
88 |
+
else:
|
89 |
+
model = hinglish_model
|
90 |
+
tokenizer = hinglish_tokenizer
|
91 |
+
labels = ["not toxic", "toxic"]
|
92 |
+
|
93 |
+
# Tokenization
|
94 |
+
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(device)
|
95 |
+
with torch.no_grad():
|
96 |
+
outputs = model(**inputs)
|
97 |
+
probs = torch.softmax(outputs, dim=1).squeeze().tolist()
|
98 |
+
|
99 |
+
response = {
|
100 |
+
"language": "english" if lang == "en" else "hinglish",
|
101 |
+
"prediction": {label: float(round(prob, 4)) for label, prob in zip(labels, probs)}
|
102 |
+
}
|
103 |
+
return response
|