Spaces:
Sleeping
Sleeping
imrane ait ouaouch commited on
Update app.py
Browse files
app.py
CHANGED
|
@@ -2,15 +2,17 @@ from fastapi import FastAPI
|
|
| 2 |
from pydantic import BaseModel
|
| 3 |
from transformers import BertTokenizer, BertForSequenceClassification
|
| 4 |
import torch
|
|
|
|
| 5 |
|
| 6 |
app = FastAPI()
|
| 7 |
|
|
|
|
| 8 |
MODEL_NAME = "ASTRAELIV/bert-spam-domain-final"
|
| 9 |
|
| 10 |
tokenizer = BertTokenizer.from_pretrained(MODEL_NAME)
|
| 11 |
model = BertForSequenceClassification.from_pretrained(MODEL_NAME)
|
| 12 |
|
| 13 |
-
model.eval()
|
| 14 |
|
| 15 |
|
| 16 |
class EmailRequest(BaseModel):
|
|
@@ -19,14 +21,23 @@ class EmailRequest(BaseModel):
|
|
| 19 |
domain: str
|
| 20 |
|
| 21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
@app.post("/predict")
|
| 23 |
def predict(request: EmailRequest):
|
| 24 |
|
|
|
|
| 25 |
text = (
|
| 26 |
f"[DOMAIN] {request.domain} "
|
| 27 |
f"[BODY] {request.subject} {request.body}"
|
| 28 |
)
|
| 29 |
|
|
|
|
| 30 |
inputs = tokenizer(
|
| 31 |
text,
|
| 32 |
return_tensors="pt",
|
|
@@ -35,19 +46,76 @@ def predict(request: EmailRequest):
|
|
| 35 |
max_length=512
|
| 36 |
)
|
| 37 |
|
| 38 |
-
|
| 39 |
-
outputs = model(**inputs)
|
| 40 |
-
probs = torch.nn.functional.softmax(outputs.logits, dim=1)
|
| 41 |
|
| 42 |
-
prediction = torch.argmax(probs, dim=1).item()
|
| 43 |
-
confidence = probs[0][prediction].item()
|
| 44 |
|
| 45 |
-
|
| 46 |
-
"prediction": "spam" if prediction == 1 else "ham",
|
| 47 |
-
"confidence": round(confidence * 100, 2)
|
| 48 |
-
}
|
| 49 |
|
|
|
|
|
|
|
| 50 |
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
from pydantic import BaseModel
|
| 3 |
from transformers import BertTokenizer, BertForSequenceClassification
|
| 4 |
import torch
|
| 5 |
+
import torch.nn.functional as F
|
| 6 |
|
| 7 |
app = FastAPI()
|
| 8 |
|
| 9 |
+
|
| 10 |
MODEL_NAME = "ASTRAELIV/bert-spam-domain-final"
|
| 11 |
|
| 12 |
tokenizer = BertTokenizer.from_pretrained(MODEL_NAME)
|
| 13 |
model = BertForSequenceClassification.from_pretrained(MODEL_NAME)
|
| 14 |
|
| 15 |
+
model.eval()
|
| 16 |
|
| 17 |
|
| 18 |
class EmailRequest(BaseModel):
|
|
|
|
| 21 |
domain: str
|
| 22 |
|
| 23 |
|
| 24 |
+
|
| 25 |
+
@app.get("/")
|
| 26 |
+
def health_check():
|
| 27 |
+
return {"status": "API is running ✅"}
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
|
| 31 |
@app.post("/predict")
|
| 32 |
def predict(request: EmailRequest):
|
| 33 |
|
| 34 |
+
|
| 35 |
text = (
|
| 36 |
f"[DOMAIN] {request.domain} "
|
| 37 |
f"[BODY] {request.subject} {request.body}"
|
| 38 |
)
|
| 39 |
|
| 40 |
+
|
| 41 |
inputs = tokenizer(
|
| 42 |
text,
|
| 43 |
return_tensors="pt",
|
|
|
|
| 46 |
max_length=512
|
| 47 |
)
|
| 48 |
|
| 49 |
+
input_ids = inputs["input_ids"]
|
|
|
|
|
|
|
| 50 |
|
|
|
|
|
|
|
| 51 |
|
| 52 |
+
torch.set_grad_enabled(True)
|
|
|
|
|
|
|
|
|
|
| 53 |
|
| 54 |
+
outputs = model(**inputs)
|
| 55 |
+
logits = outputs.logits
|
| 56 |
|
| 57 |
+
|
| 58 |
+
probs = F.softmax(logits, dim=1)
|
| 59 |
+
|
| 60 |
+
prediction = torch.argmax(probs, dim=1).item()
|
| 61 |
+
confidence = probs[0][prediction].item()
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
model.zero_grad()
|
| 65 |
+
logits[0][prediction].backward()
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
embeddings = model.bert.embeddings.word_embeddings
|
| 69 |
+
grads = embeddings.weight.grad
|
| 70 |
+
|
| 71 |
+
tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
|
| 72 |
+
|
| 73 |
+
token_scores = []
|
| 74 |
+
|
| 75 |
+
for token_id, token in zip(input_ids[0], tokens):
|
| 76 |
+
grad = grads[token_id]
|
| 77 |
+
score = torch.norm(grad).item()
|
| 78 |
+
token_scores.append((token, score))
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
max_score = max(score for _, score in token_scores) + 1e-8
|
| 82 |
+
|
| 83 |
+
token_scores = [
|
| 84 |
+
(tok, score / max_score)
|
| 85 |
+
for tok, score in token_scores
|
| 86 |
+
if tok not in ["[CLS]", "[SEP]", "[PAD]"]
|
| 87 |
+
]
|
| 88 |
+
|
| 89 |
+
sorted_tokens = sorted(
|
| 90 |
+
token_scores,
|
| 91 |
+
key=lambda x: x[1],
|
| 92 |
+
reverse=True
|
| 93 |
+
)
|
| 94 |
+
|
| 95 |
+
top_tokens = sorted_tokens[:8]
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
torch.set_grad_enabled(False)
|
| 99 |
+
|
| 100 |
+
return {
|
| 101 |
+
"prediction": "spam" if prediction == 1 else "ham",
|
| 102 |
+
"confidence": round(confidence * 100, 2),
|
| 103 |
+
"probabilities": {
|
| 104 |
+
"spam": round(probs[0][1].item() * 100, 2),
|
| 105 |
+
"ham": round(probs[0][0].item() * 100, 2)
|
| 106 |
+
},
|
| 107 |
+
"top_indicators": [
|
| 108 |
+
{
|
| 109 |
+
"token": tok,
|
| 110 |
+
"score": round(score, 3)
|
| 111 |
+
}
|
| 112 |
+
for tok, score in top_tokens
|
| 113 |
+
],
|
| 114 |
+
"token_scores": [
|
| 115 |
+
{
|
| 116 |
+
"token": tok,
|
| 117 |
+
"score": round(score, 3)
|
| 118 |
+
}
|
| 119 |
+
for tok, score in token_scores
|
| 120 |
+
]
|
| 121 |
+
}
|