Spaces:
Runtime error
Runtime error
feat: add highlight
Browse files- main.py +83 -18
- templates/index.html +26 -0
main.py
CHANGED
|
@@ -77,6 +77,69 @@ class PredictionResponse(BaseModel):
|
|
| 77 |
probabilities: dict # Probabilities for each class
|
| 78 |
confidence: float # Probability for the predicted class
|
| 79 |
status: str # "judi", "bukan judi", or "ragu-ragu"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
|
| 81 |
@app.get("/")
|
| 82 |
def read_root(request: Request):
|
|
@@ -87,18 +150,17 @@ def read_root(request: Request):
|
|
| 87 |
@app.post("/predict", response_model=PredictionResponse)
|
| 88 |
async def predict(input_data: PredictionInput):
|
| 89 |
try:
|
| 90 |
-
# Threshold values
|
| 91 |
threshold_low = 0.4
|
| 92 |
threshold_high = 0.7
|
| 93 |
-
|
| 94 |
-
text = str(input_data.text) if input_data.text
|
| 95 |
-
|
| 96 |
# Tokenize input text
|
| 97 |
inputs = tokenizer(
|
| 98 |
-
text,
|
| 99 |
-
return_tensors="pt",
|
| 100 |
-
padding=True,
|
| 101 |
-
truncation=True,
|
| 102 |
max_length=128
|
| 103 |
)
|
| 104 |
inputs = {key: val.to(device) for key, val in inputs.items()}
|
|
@@ -107,34 +169,37 @@ async def predict(input_data: PredictionInput):
|
|
| 107 |
with torch.no_grad():
|
| 108 |
outputs = model(**inputs)
|
| 109 |
probs = F.softmax(outputs.logits, dim=-1)
|
| 110 |
-
prob_0 = probs[0][0].item()
|
| 111 |
-
prob_1 = probs[0][1].item()
|
| 112 |
-
|
| 113 |
-
#
|
| 114 |
if prob_1 >= threshold_high:
|
| 115 |
-
prediction = 1
|
| 116 |
prediction_text = f'judi ({prob_1:.2f})'
|
| 117 |
status = "judi"
|
| 118 |
confidence = prob_1
|
| 119 |
elif prob_1 <= threshold_low:
|
| 120 |
-
prediction = 0
|
| 121 |
prediction_text = f'bukan judi ({prob_0:.2f})'
|
| 122 |
status = "bukan judi"
|
| 123 |
confidence = prob_0
|
| 124 |
else:
|
| 125 |
-
prediction = -1
|
| 126 |
prediction_text = f'ragu-ragu ({prob_1:.2f})'
|
| 127 |
status = "ragu-ragu"
|
| 128 |
confidence = max(prob_0, prob_1)
|
| 129 |
-
|
|
|
|
|
|
|
| 130 |
return PredictionResponse(
|
| 131 |
prediction=prediction,
|
| 132 |
prediction_text=prediction_text,
|
| 133 |
probabilities={"prob_bukan_judi": prob_0, "prob_judi": prob_1},
|
| 134 |
confidence=confidence,
|
| 135 |
-
status=status
|
|
|
|
| 136 |
)
|
| 137 |
-
|
| 138 |
except Exception as e:
|
| 139 |
raise HTTPException(status_code=400, detail=f"Prediction error: {str(e)}")
|
| 140 |
|
|
|
|
| 77 |
probabilities: dict # Probabilities for each class
|
| 78 |
confidence: float # Probability for the predicted class
|
| 79 |
status: str # "judi", "bukan judi", or "ragu-ragu"
|
| 80 |
+
highlighted_tokens: list[str]
|
| 81 |
+
|
| 82 |
+
def get_important_tokens(text: str, tokenizer, model, target_class=1, top_k=3):
|
| 83 |
+
model.eval()
|
| 84 |
+
|
| 85 |
+
# Tokenize dengan offset mapping untuk kata asli
|
| 86 |
+
encoding = tokenizer(
|
| 87 |
+
text,
|
| 88 |
+
return_tensors="pt",
|
| 89 |
+
return_offsets_mapping=True,
|
| 90 |
+
padding=True,
|
| 91 |
+
truncation=True,
|
| 92 |
+
max_length=128
|
| 93 |
+
)
|
| 94 |
+
input_ids = encoding["input_ids"].to(model.device)
|
| 95 |
+
attention_mask = encoding["attention_mask"].to(model.device)
|
| 96 |
+
offset_mapping = encoding["offset_mapping"][0].tolist()
|
| 97 |
+
|
| 98 |
+
# Dapatkan embeddings
|
| 99 |
+
embedding_layer = model.get_input_embeddings()
|
| 100 |
+
input_embeddings = embedding_layer(input_ids)
|
| 101 |
+
input_embeddings.retain_grad()
|
| 102 |
+
|
| 103 |
+
# Forward pass
|
| 104 |
+
outputs = model(inputs_embeds=input_embeddings, attention_mask=attention_mask)
|
| 105 |
+
probs = F.softmax(outputs.logits, dim=-1)
|
| 106 |
+
target_score = probs[0][target_class]
|
| 107 |
+
|
| 108 |
+
# Backward pass
|
| 109 |
+
model.zero_grad()
|
| 110 |
+
target_score.backward()
|
| 111 |
+
|
| 112 |
+
# Importance score per token
|
| 113 |
+
grads = input_embeddings.grad.abs().sum(dim=-1).squeeze().tolist()
|
| 114 |
+
|
| 115 |
+
# Mapping: offset range → importance score
|
| 116 |
+
token_scores = []
|
| 117 |
+
for i, (start, end) in enumerate(offset_mapping):
|
| 118 |
+
if start == end: # Special tokens
|
| 119 |
+
continue
|
| 120 |
+
token_scores.append((start, end, grads[i]))
|
| 121 |
+
|
| 122 |
+
# Gabungkan berdasarkan overlap character range
|
| 123 |
+
from collections import defaultdict
|
| 124 |
+
char_score_map = defaultdict(float)
|
| 125 |
+
for start, end, score in token_scores:
|
| 126 |
+
for i in range(start, end):
|
| 127 |
+
char_score_map[i] += score
|
| 128 |
+
|
| 129 |
+
# Bagi teks ke kata
|
| 130 |
+
import re
|
| 131 |
+
word_spans = [(m.start(), m.end(), m.group()) for m in re.finditer(r'\S+', text)]
|
| 132 |
+
|
| 133 |
+
word_scores = []
|
| 134 |
+
for start, end, word in word_spans:
|
| 135 |
+
score = sum(char_score_map[i] for i in range(start, end))
|
| 136 |
+
word_scores.append((word, score))
|
| 137 |
+
|
| 138 |
+
# Ambil kata dengan skor tertinggi
|
| 139 |
+
word_scores.sort(key=lambda x: x[1], reverse=True)
|
| 140 |
+
highlighted_words = [w for w, s in word_scores[:top_k]]
|
| 141 |
+
|
| 142 |
+
return highlighted_words
|
| 143 |
|
| 144 |
@app.get("/")
|
| 145 |
def read_root(request: Request):
|
|
|
|
| 150 |
@app.post("/predict", response_model=PredictionResponse)
|
| 151 |
async def predict(input_data: PredictionInput):
|
| 152 |
try:
|
|
|
|
| 153 |
threshold_low = 0.4
|
| 154 |
threshold_high = 0.7
|
| 155 |
+
|
| 156 |
+
text = str(input_data.text) if input_data.text else ""
|
| 157 |
+
|
| 158 |
# Tokenize input text
|
| 159 |
inputs = tokenizer(
|
| 160 |
+
text,
|
| 161 |
+
return_tensors="pt",
|
| 162 |
+
padding=True,
|
| 163 |
+
truncation=True,
|
| 164 |
max_length=128
|
| 165 |
)
|
| 166 |
inputs = {key: val.to(device) for key, val in inputs.items()}
|
|
|
|
| 169 |
with torch.no_grad():
|
| 170 |
outputs = model(**inputs)
|
| 171 |
probs = F.softmax(outputs.logits, dim=-1)
|
| 172 |
+
prob_0 = probs[0][0].item()
|
| 173 |
+
prob_1 = probs[0][1].item()
|
| 174 |
+
|
| 175 |
+
# Interpret prediction
|
| 176 |
if prob_1 >= threshold_high:
|
| 177 |
+
prediction = 1
|
| 178 |
prediction_text = f'judi ({prob_1:.2f})'
|
| 179 |
status = "judi"
|
| 180 |
confidence = prob_1
|
| 181 |
elif prob_1 <= threshold_low:
|
| 182 |
+
prediction = 0
|
| 183 |
prediction_text = f'bukan judi ({prob_0:.2f})'
|
| 184 |
status = "bukan judi"
|
| 185 |
confidence = prob_0
|
| 186 |
else:
|
| 187 |
+
prediction = -1
|
| 188 |
prediction_text = f'ragu-ragu ({prob_1:.2f})'
|
| 189 |
status = "ragu-ragu"
|
| 190 |
confidence = max(prob_0, prob_1)
|
| 191 |
+
|
| 192 |
+
highlighted_tokens = get_important_tokens(text, tokenizer, model, target_class=1)
|
| 193 |
+
|
| 194 |
return PredictionResponse(
|
| 195 |
prediction=prediction,
|
| 196 |
prediction_text=prediction_text,
|
| 197 |
probabilities={"prob_bukan_judi": prob_0, "prob_judi": prob_1},
|
| 198 |
confidence=confidence,
|
| 199 |
+
status=status,
|
| 200 |
+
highlighted_tokens=highlighted_tokens
|
| 201 |
)
|
| 202 |
+
|
| 203 |
except Exception as e:
|
| 204 |
raise HTTPException(status_code=400, detail=f"Prediction error: {str(e)}")
|
| 205 |
|
templates/index.html
CHANGED
|
@@ -53,6 +53,15 @@
|
|
| 53 |
.gradient-error {
|
| 54 |
background: linear-gradient(135deg, #ff6b6b 0%, #ffa500 100%);
|
| 55 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
</style>
|
| 57 |
</head>
|
| 58 |
|
|
@@ -203,6 +212,16 @@
|
|
| 203 |
statusExplanation = 'Model tidak yakin dengan klasifikasi teks ini';
|
| 204 |
}
|
| 205 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 206 |
resultContent.innerHTML = `
|
| 207 |
<div class="space-y-2 sm:space-y-3">
|
| 208 |
<p><strong>Status:</strong> ${data.status}</p>
|
|
@@ -245,6 +264,13 @@
|
|
| 245 |
</div>
|
| 246 |
`;
|
| 247 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 248 |
result.classList.remove('hidden');
|
| 249 |
}
|
| 250 |
|
|
|
|
| 53 |
.gradient-error {
|
| 54 |
background: linear-gradient(135deg, #ff6b6b 0%, #ffa500 100%);
|
| 55 |
}
|
| 56 |
+
|
| 57 |
+
.highlight-word {
|
| 58 |
+
background-color: rgba(255, 255, 255, 0.85);
|
| 59 |
+
color: #d6336c;
|
| 60 |
+
font-weight: 600;
|
| 61 |
+
padding: 0 4px;
|
| 62 |
+
border-radius: 4px;
|
| 63 |
+
box-shadow: 0 1px 2px rgba(0,0,0,0.1);
|
| 64 |
+
}
|
| 65 |
</style>
|
| 66 |
</head>
|
| 67 |
|
|
|
|
| 212 |
statusExplanation = 'Model tidak yakin dengan klasifikasi teks ini';
|
| 213 |
}
|
| 214 |
|
| 215 |
+
const originalText = document.getElementById('textInput').value;
|
| 216 |
+
const importantWords = data.prediction === 1 ? data.highlighted_tokens || [] : [];
|
| 217 |
+
|
| 218 |
+
let highlightedText = originalText;
|
| 219 |
+
|
| 220 |
+
importantWords.forEach(word => {
|
| 221 |
+
const regex = new RegExp(`(${word})`, 'gi');
|
| 222 |
+
highlightedText = highlightedText.replace(regex, '<span class="highlight-word">$1</span>');
|
| 223 |
+
});
|
| 224 |
+
|
| 225 |
resultContent.innerHTML = `
|
| 226 |
<div class="space-y-2 sm:space-y-3">
|
| 227 |
<p><strong>Status:</strong> ${data.status}</p>
|
|
|
|
| 264 |
</div>
|
| 265 |
`;
|
| 266 |
|
| 267 |
+
resultContent.innerHTML += `
|
| 268 |
+
<div class="mt-4 sm:mt-6 bg-white bg-opacity-20 p-4 rounded text-xs sm:text-sm">
|
| 269 |
+
<strong>Highlight Kata Penting:</strong>
|
| 270 |
+
<p class="mt-2">${highlightedText}</p>
|
| 271 |
+
</div>
|
| 272 |
+
`;
|
| 273 |
+
|
| 274 |
result.classList.remove('hidden');
|
| 275 |
}
|
| 276 |
|