Update app.py
Browse files
app.py
CHANGED
@@ -13,18 +13,18 @@ from functools import partial
|
|
13 |
import time
|
14 |
from datetime import datetime
|
15 |
|
16 |
-
|
17 |
logging.basicConfig(level=logging.INFO)
|
18 |
logger = logging.getLogger(__name__)
|
19 |
|
20 |
-
|
21 |
MAX_LENGTH = 512
|
22 |
MODEL_NAME = "microsoft/deberta-v3-small"
|
23 |
WINDOW_SIZE = 6
|
24 |
WINDOW_OVERLAP = 2
|
25 |
CONFIDENCE_THRESHOLD = 0.65
|
26 |
-
BATCH_SIZE = 8
|
27 |
-
MAX_WORKERS = 4
|
28 |
|
29 |
class TextWindowProcessor:
|
30 |
def __init__(self):
|
@@ -41,7 +41,7 @@ class TextWindowProcessor:
|
|
41 |
disabled_pipes = [pipe for pipe in self.nlp.pipe_names if pipe != 'sentencizer']
|
42 |
self.nlp.disable_pipes(*disabled_pipes)
|
43 |
|
44 |
-
|
45 |
self.executor = ThreadPoolExecutor(max_workers=MAX_WORKERS)
|
46 |
|
47 |
def split_into_sentences(self, text: str) -> List[str]:
|
@@ -60,17 +60,16 @@ class TextWindowProcessor:
|
|
60 |
return windows
|
61 |
|
62 |
def create_centered_windows(self, sentences: List[str], window_size: int) -> Tuple[List[str], List[List[int]]]:
|
63 |
-
"""Create windows with better boundary handling"""
|
64 |
windows = []
|
65 |
window_sentence_indices = []
|
66 |
|
67 |
for i in range(len(sentences)):
|
68 |
-
|
69 |
half_window = window_size // 2
|
70 |
start_idx = max(0, i - half_window)
|
71 |
end_idx = min(len(sentences), i + half_window + 1)
|
72 |
|
73 |
-
|
74 |
window = sentences[start_idx:end_idx]
|
75 |
windows.append(" ".join(window))
|
76 |
window_sentence_indices.append(list(range(start_idx, end_idx)))
|
@@ -79,7 +78,7 @@ class TextWindowProcessor:
|
|
79 |
|
80 |
class TextClassifier:
|
81 |
def __init__(self):
|
82 |
-
|
83 |
if not torch.cuda.is_available():
|
84 |
torch.set_num_threads(MAX_WORKERS)
|
85 |
torch.set_num_interop_threads(MAX_WORKERS)
|
@@ -119,7 +118,6 @@ class TextClassifier:
|
|
119 |
self.model.eval()
|
120 |
|
121 |
def quick_scan(self, text: str) -> Dict:
|
122 |
-
"""Perform a quick scan using simple window analysis."""
|
123 |
if not text.strip():
|
124 |
return {
|
125 |
'prediction': 'unknown',
|
@@ -132,7 +130,7 @@ class TextClassifier:
|
|
132 |
|
133 |
predictions = []
|
134 |
|
135 |
-
|
136 |
for i in range(0, len(windows), BATCH_SIZE):
|
137 |
batch_windows = windows[i:i + BATCH_SIZE]
|
138 |
|
@@ -157,7 +155,7 @@ class TextClassifier:
|
|
157 |
}
|
158 |
predictions.append(prediction)
|
159 |
|
160 |
-
|
161 |
del inputs, outputs, probs
|
162 |
if torch.cuda.is_available():
|
163 |
torch.cuda.empty_cache()
|
@@ -179,8 +177,7 @@ class TextClassifier:
|
|
179 |
}
|
180 |
|
181 |
def detailed_scan(self, text: str) -> Dict:
|
182 |
-
|
183 |
-
# Clean up trailing whitespace
|
184 |
text = text.rstrip()
|
185 |
|
186 |
if not text.strip():
|
@@ -199,14 +196,14 @@ class TextClassifier:
|
|
199 |
if not sentences:
|
200 |
return {}
|
201 |
|
202 |
-
|
203 |
windows, window_sentence_indices = self.processor.create_centered_windows(sentences, WINDOW_SIZE)
|
204 |
|
205 |
-
|
206 |
sentence_appearances = {i: 0 for i in range(len(sentences))}
|
207 |
sentence_scores = {i: {'human_prob': 0.0, 'ai_prob': 0.0} for i in range(len(sentences))}
|
208 |
|
209 |
-
|
210 |
for i in range(0, len(windows), BATCH_SIZE):
|
211 |
batch_windows = windows[i:i + BATCH_SIZE]
|
212 |
batch_indices = window_sentence_indices[i:i + BATCH_SIZE]
|
@@ -223,45 +220,45 @@ class TextClassifier:
|
|
223 |
outputs = self.model(**inputs)
|
224 |
probs = F.softmax(outputs.logits, dim=-1)
|
225 |
|
226 |
-
|
227 |
for window_idx, indices in enumerate(batch_indices):
|
228 |
center_idx = len(indices) // 2
|
229 |
-
center_weight = 0.7
|
230 |
-
edge_weight = 0.3 / (len(indices) - 1)
|
231 |
|
232 |
for pos, sent_idx in enumerate(indices):
|
233 |
-
|
234 |
weight = center_weight if pos == center_idx else edge_weight
|
235 |
sentence_appearances[sent_idx] += weight
|
236 |
sentence_scores[sent_idx]['human_prob'] += weight * probs[window_idx][1].item()
|
237 |
sentence_scores[sent_idx]['ai_prob'] += weight * probs[window_idx][0].item()
|
238 |
|
239 |
-
|
240 |
del inputs, outputs, probs
|
241 |
if torch.cuda.is_available():
|
242 |
torch.cuda.empty_cache()
|
243 |
|
244 |
-
|
245 |
sentence_predictions = []
|
246 |
for i in range(len(sentences)):
|
247 |
if sentence_appearances[i] > 0:
|
248 |
human_prob = sentence_scores[i]['human_prob'] / sentence_appearances[i]
|
249 |
ai_prob = sentence_scores[i]['ai_prob'] / sentence_appearances[i]
|
250 |
|
251 |
-
|
252 |
if i > 0 and i < len(sentences) - 1:
|
253 |
prev_human = sentence_scores[i-1]['human_prob'] / sentence_appearances[i-1]
|
254 |
prev_ai = sentence_scores[i-1]['ai_prob'] / sentence_appearances[i-1]
|
255 |
next_human = sentence_scores[i+1]['human_prob'] / sentence_appearances[i+1]
|
256 |
next_ai = sentence_scores[i+1]['ai_prob'] / sentence_appearances[i+1]
|
257 |
|
258 |
-
|
259 |
current_pred = 'human' if human_prob > ai_prob else 'ai'
|
260 |
prev_pred = 'human' if prev_human > prev_ai else 'ai'
|
261 |
next_pred = 'human' if next_human > next_ai else 'ai'
|
262 |
|
263 |
if current_pred != prev_pred or current_pred != next_pred:
|
264 |
-
|
265 |
smooth_factor = 0.1
|
266 |
human_prob = (human_prob * (1 - smooth_factor) +
|
267 |
(prev_human + next_human) * smooth_factor / 2)
|
@@ -284,7 +281,6 @@ class TextClassifier:
|
|
284 |
}
|
285 |
|
286 |
def format_predictions_html(self, sentence_predictions: List[Dict]) -> str:
|
287 |
-
"""Format predictions as HTML with color-coding."""
|
288 |
html_parts = []
|
289 |
|
290 |
for pred in sentence_predictions:
|
@@ -293,21 +289,20 @@ class TextClassifier:
|
|
293 |
|
294 |
if confidence >= CONFIDENCE_THRESHOLD:
|
295 |
if pred['prediction'] == 'human':
|
296 |
-
color = "
|
297 |
else:
|
298 |
-
color = "
|
299 |
else:
|
300 |
if pred['prediction'] == 'human':
|
301 |
-
color = "
|
302 |
else:
|
303 |
-
color = "
|
304 |
|
305 |
html_parts.append(f'<span style="background-color: {color};">{sentence}</span>')
|
306 |
|
307 |
return " ".join(html_parts)
|
308 |
|
309 |
def aggregate_predictions(self, predictions: List[Dict]) -> Dict:
|
310 |
-
"""Aggregate predictions from multiple sentences into a single prediction."""
|
311 |
if not predictions:
|
312 |
return {
|
313 |
'prediction': 'unknown',
|
@@ -329,14 +324,13 @@ class TextClassifier:
|
|
329 |
}
|
330 |
|
331 |
def analyze_text(text: str, mode: str, classifier: TextClassifier) -> tuple:
|
332 |
-
|
333 |
-
# Start timing for normal analysis
|
334 |
start_time = time.time()
|
335 |
|
336 |
-
|
337 |
word_count = len(text.split())
|
338 |
|
339 |
-
|
340 |
original_mode = mode
|
341 |
if word_count < 200 and mode == "detailed":
|
342 |
mode = "quick"
|
@@ -350,15 +344,15 @@ def analyze_text(text: str, mode: str, classifier: TextClassifier) -> tuple:
|
|
350 |
Windows analyzed: {result['num_windows']}
|
351 |
"""
|
352 |
|
353 |
-
|
354 |
if original_mode == "detailed":
|
355 |
quick_analysis += f"\n\nNote: Switched to quick mode because text contains only {word_count} words. Minimum 200 words required for detailed analysis."
|
356 |
|
357 |
-
|
358 |
execution_time = (time.time() - start_time) * 1000
|
359 |
|
360 |
return (
|
361 |
-
text,
|
362 |
"Quick scan mode - no sentence-level analysis available",
|
363 |
quick_analysis
|
364 |
)
|
@@ -380,7 +374,7 @@ def analyze_text(text: str, mode: str, classifier: TextClassifier) -> tuple:
|
|
380 |
Number of sentences analyzed: {final_pred['num_sentences']}
|
381 |
"""
|
382 |
|
383 |
-
|
384 |
execution_time = (time.time() - start_time) * 1000
|
385 |
|
386 |
return (
|
@@ -389,10 +383,10 @@ def analyze_text(text: str, mode: str, classifier: TextClassifier) -> tuple:
|
|
389 |
overall_result
|
390 |
)
|
391 |
|
392 |
-
|
393 |
classifier = TextClassifier()
|
394 |
|
395 |
-
|
396 |
demo = gr.Interface(
|
397 |
fn=lambda text, mode: analyze_text(text, mode, classifier),
|
398 |
inputs=[
|
@@ -419,19 +413,17 @@ demo = gr.Interface(
|
|
419 |
flagging_mode="never"
|
420 |
)
|
421 |
|
422 |
-
|
423 |
app = demo.app
|
424 |
|
425 |
-
# Add CORS middleware
|
426 |
app.add_middleware(
|
427 |
CORSMiddleware,
|
428 |
-
allow_origins=["*"],
|
429 |
allow_credentials=True,
|
430 |
allow_methods=["GET", "POST", "OPTIONS"],
|
431 |
allow_headers=["*"],
|
432 |
)
|
433 |
|
434 |
-
# Ensure CORS is applied before launching
|
435 |
if __name__ == "__main__":
|
436 |
demo.queue()
|
437 |
demo.launch(
|
|
|
13 |
import time
|
14 |
from datetime import datetime
|
15 |
|
16 |
+
|
17 |
logging.basicConfig(level=logging.INFO)
|
18 |
logger = logging.getLogger(__name__)
|
19 |
|
20 |
+
|
21 |
MAX_LENGTH = 512
|
22 |
MODEL_NAME = "microsoft/deberta-v3-small"
|
23 |
WINDOW_SIZE = 6
|
24 |
WINDOW_OVERLAP = 2
|
25 |
CONFIDENCE_THRESHOLD = 0.65
|
26 |
+
BATCH_SIZE = 8
|
27 |
+
MAX_WORKERS = 4
|
28 |
|
29 |
class TextWindowProcessor:
|
30 |
def __init__(self):
|
|
|
41 |
disabled_pipes = [pipe for pipe in self.nlp.pipe_names if pipe != 'sentencizer']
|
42 |
self.nlp.disable_pipes(*disabled_pipes)
|
43 |
|
44 |
+
|
45 |
self.executor = ThreadPoolExecutor(max_workers=MAX_WORKERS)
|
46 |
|
47 |
def split_into_sentences(self, text: str) -> List[str]:
|
|
|
60 |
return windows
|
61 |
|
62 |
def create_centered_windows(self, sentences: List[str], window_size: int) -> Tuple[List[str], List[List[int]]]:
|
|
|
63 |
windows = []
|
64 |
window_sentence_indices = []
|
65 |
|
66 |
for i in range(len(sentences)):
|
67 |
+
|
68 |
half_window = window_size // 2
|
69 |
start_idx = max(0, i - half_window)
|
70 |
end_idx = min(len(sentences), i + half_window + 1)
|
71 |
|
72 |
+
|
73 |
window = sentences[start_idx:end_idx]
|
74 |
windows.append(" ".join(window))
|
75 |
window_sentence_indices.append(list(range(start_idx, end_idx)))
|
|
|
78 |
|
79 |
class TextClassifier:
|
80 |
def __init__(self):
|
81 |
+
|
82 |
if not torch.cuda.is_available():
|
83 |
torch.set_num_threads(MAX_WORKERS)
|
84 |
torch.set_num_interop_threads(MAX_WORKERS)
|
|
|
118 |
self.model.eval()
|
119 |
|
120 |
def quick_scan(self, text: str) -> Dict:
|
|
|
121 |
if not text.strip():
|
122 |
return {
|
123 |
'prediction': 'unknown',
|
|
|
130 |
|
131 |
predictions = []
|
132 |
|
133 |
+
|
134 |
for i in range(0, len(windows), BATCH_SIZE):
|
135 |
batch_windows = windows[i:i + BATCH_SIZE]
|
136 |
|
|
|
155 |
}
|
156 |
predictions.append(prediction)
|
157 |
|
158 |
+
|
159 |
del inputs, outputs, probs
|
160 |
if torch.cuda.is_available():
|
161 |
torch.cuda.empty_cache()
|
|
|
177 |
}
|
178 |
|
179 |
def detailed_scan(self, text: str) -> Dict:
|
180 |
+
|
|
|
181 |
text = text.rstrip()
|
182 |
|
183 |
if not text.strip():
|
|
|
196 |
if not sentences:
|
197 |
return {}
|
198 |
|
199 |
+
|
200 |
windows, window_sentence_indices = self.processor.create_centered_windows(sentences, WINDOW_SIZE)
|
201 |
|
202 |
+
|
203 |
sentence_appearances = {i: 0 for i in range(len(sentences))}
|
204 |
sentence_scores = {i: {'human_prob': 0.0, 'ai_prob': 0.0} for i in range(len(sentences))}
|
205 |
|
206 |
+
|
207 |
for i in range(0, len(windows), BATCH_SIZE):
|
208 |
batch_windows = windows[i:i + BATCH_SIZE]
|
209 |
batch_indices = window_sentence_indices[i:i + BATCH_SIZE]
|
|
|
220 |
outputs = self.model(**inputs)
|
221 |
probs = F.softmax(outputs.logits, dim=-1)
|
222 |
|
223 |
+
|
224 |
for window_idx, indices in enumerate(batch_indices):
|
225 |
center_idx = len(indices) // 2
|
226 |
+
center_weight = 0.7
|
227 |
+
edge_weight = 0.3 / (len(indices) - 1)
|
228 |
|
229 |
for pos, sent_idx in enumerate(indices):
|
230 |
+
|
231 |
weight = center_weight if pos == center_idx else edge_weight
|
232 |
sentence_appearances[sent_idx] += weight
|
233 |
sentence_scores[sent_idx]['human_prob'] += weight * probs[window_idx][1].item()
|
234 |
sentence_scores[sent_idx]['ai_prob'] += weight * probs[window_idx][0].item()
|
235 |
|
236 |
+
|
237 |
del inputs, outputs, probs
|
238 |
if torch.cuda.is_available():
|
239 |
torch.cuda.empty_cache()
|
240 |
|
241 |
+
|
242 |
sentence_predictions = []
|
243 |
for i in range(len(sentences)):
|
244 |
if sentence_appearances[i] > 0:
|
245 |
human_prob = sentence_scores[i]['human_prob'] / sentence_appearances[i]
|
246 |
ai_prob = sentence_scores[i]['ai_prob'] / sentence_appearances[i]
|
247 |
|
248 |
+
|
249 |
if i > 0 and i < len(sentences) - 1:
|
250 |
prev_human = sentence_scores[i-1]['human_prob'] / sentence_appearances[i-1]
|
251 |
prev_ai = sentence_scores[i-1]['ai_prob'] / sentence_appearances[i-1]
|
252 |
next_human = sentence_scores[i+1]['human_prob'] / sentence_appearances[i+1]
|
253 |
next_ai = sentence_scores[i+1]['ai_prob'] / sentence_appearances[i+1]
|
254 |
|
255 |
+
|
256 |
current_pred = 'human' if human_prob > ai_prob else 'ai'
|
257 |
prev_pred = 'human' if prev_human > prev_ai else 'ai'
|
258 |
next_pred = 'human' if next_human > next_ai else 'ai'
|
259 |
|
260 |
if current_pred != prev_pred or current_pred != next_pred:
|
261 |
+
|
262 |
smooth_factor = 0.1
|
263 |
human_prob = (human_prob * (1 - smooth_factor) +
|
264 |
(prev_human + next_human) * smooth_factor / 2)
|
|
|
281 |
}
|
282 |
|
283 |
def format_predictions_html(self, sentence_predictions: List[Dict]) -> str:
|
|
|
284 |
html_parts = []
|
285 |
|
286 |
for pred in sentence_predictions:
|
|
|
289 |
|
290 |
if confidence >= CONFIDENCE_THRESHOLD:
|
291 |
if pred['prediction'] == 'human':
|
292 |
+
color = "
|
293 |
else:
|
294 |
+
color = "
|
295 |
else:
|
296 |
if pred['prediction'] == 'human':
|
297 |
+
color = "
|
298 |
else:
|
299 |
+
color = "
|
300 |
|
301 |
html_parts.append(f'<span style="background-color: {color};">{sentence}</span>')
|
302 |
|
303 |
return " ".join(html_parts)
|
304 |
|
305 |
def aggregate_predictions(self, predictions: List[Dict]) -> Dict:
|
|
|
306 |
if not predictions:
|
307 |
return {
|
308 |
'prediction': 'unknown',
|
|
|
324 |
}
|
325 |
|
326 |
def analyze_text(text: str, mode: str, classifier: TextClassifier) -> tuple:
|
327 |
+
|
|
|
328 |
start_time = time.time()
|
329 |
|
330 |
+
|
331 |
word_count = len(text.split())
|
332 |
|
333 |
+
|
334 |
original_mode = mode
|
335 |
if word_count < 200 and mode == "detailed":
|
336 |
mode = "quick"
|
|
|
344 |
Windows analyzed: {result['num_windows']}
|
345 |
"""
|
346 |
|
347 |
+
|
348 |
if original_mode == "detailed":
|
349 |
quick_analysis += f"\n\nNote: Switched to quick mode because text contains only {word_count} words. Minimum 200 words required for detailed analysis."
|
350 |
|
351 |
+
|
352 |
execution_time = (time.time() - start_time) * 1000
|
353 |
|
354 |
return (
|
355 |
+
text,
|
356 |
"Quick scan mode - no sentence-level analysis available",
|
357 |
quick_analysis
|
358 |
)
|
|
|
374 |
Number of sentences analyzed: {final_pred['num_sentences']}
|
375 |
"""
|
376 |
|
377 |
+
|
378 |
execution_time = (time.time() - start_time) * 1000
|
379 |
|
380 |
return (
|
|
|
383 |
overall_result
|
384 |
)
|
385 |
|
386 |
+
|
387 |
classifier = TextClassifier()
|
388 |
|
389 |
+
|
390 |
demo = gr.Interface(
|
391 |
fn=lambda text, mode: analyze_text(text, mode, classifier),
|
392 |
inputs=[
|
|
|
413 |
flagging_mode="never"
|
414 |
)
|
415 |
|
416 |
+
|
417 |
app = demo.app
|
418 |
|
|
|
419 |
app.add_middleware(
|
420 |
CORSMiddleware,
|
421 |
+
allow_origins=["*"],
|
422 |
allow_credentials=True,
|
423 |
allow_methods=["GET", "POST", "OPTIONS"],
|
424 |
allow_headers=["*"],
|
425 |
)
|
426 |
|
|
|
427 |
if __name__ == "__main__":
|
428 |
demo.queue()
|
429 |
demo.launch(
|