Update main.py
Browse files
main.py
CHANGED
@@ -2,6 +2,7 @@ from flask import Flask, request, jsonify
|
|
2 |
import torch
|
3 |
from transformers import RobertaTokenizer, RobertaForSequenceClassification
|
4 |
import os
|
|
|
5 |
from functools import lru_cache
|
6 |
|
7 |
app = Flask(__name__)
|
@@ -25,7 +26,11 @@ def load_tokenizer():
|
|
25 |
return tokenizer
|
26 |
except Exception as e:
|
27 |
print(f"Error loading tokenizer: {e}")
|
28 |
-
|
|
|
|
|
|
|
|
|
29 |
|
30 |
def load_model():
|
31 |
global device
|
@@ -60,27 +65,27 @@ def load_model():
|
|
60 |
|
61 |
except Exception as e:
|
62 |
print(f"Error loading model: {e}")
|
63 |
-
|
64 |
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
code,
|
70 |
-
truncation=True,
|
71 |
-
padding='max_length',
|
72 |
-
max_length=max_length,
|
73 |
-
return_tensors='pt'
|
74 |
-
)
|
75 |
|
76 |
try:
|
77 |
print("Loading tokenizer...")
|
78 |
tokenizer = load_tokenizer()
|
79 |
-
|
|
|
|
|
|
|
80 |
|
81 |
print("Loading model...")
|
82 |
model = load_model()
|
83 |
-
|
|
|
|
|
|
|
84 |
|
85 |
except Exception as e:
|
86 |
print(f"Error during initialization: {str(e)}")
|
@@ -111,75 +116,199 @@ def predict_batch():
|
|
111 |
codes = data['codes']
|
112 |
if not isinstance(codes, list) or len(codes) == 0:
|
113 |
return jsonify({"error": "'codes' must be a non-empty array"}), 400
|
114 |
-
|
115 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
results = []
|
117 |
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
|
127 |
return jsonify({"results": results})
|
128 |
|
129 |
except Exception as e:
|
|
|
130 |
return jsonify({"error": f"Batch prediction error: {str(e)}"}), 500
|
131 |
|
132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
def predict_vulnerability(code):
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
|
158 |
def predict_vulnerability_batch(codes):
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
codes
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
183 |
|
184 |
@app.route("/health", methods=['GET'])
|
185 |
def health_check():
|
|
|
2 |
import torch
|
3 |
from transformers import RobertaTokenizer, RobertaForSequenceClassification
|
4 |
import os
|
5 |
+
import gc
|
6 |
from functools import lru_cache
|
7 |
|
8 |
app = Flask(__name__)
|
|
|
26 |
return tokenizer
|
27 |
except Exception as e:
|
28 |
print(f"Error loading tokenizer: {e}")
|
29 |
+
try:
|
30 |
+
return RobertaTokenizer.from_pretrained('microsoft/codebert-base')
|
31 |
+
except Exception as e2:
|
32 |
+
print(f"Fallback tokenizer failed: {e2}")
|
33 |
+
return None
|
34 |
|
35 |
def load_model():
|
36 |
global device
|
|
|
65 |
|
66 |
except Exception as e:
|
67 |
print(f"Error loading model: {e}")
|
68 |
+
return None
|
69 |
|
70 |
+
def cleanup_gpu_memory():
|
71 |
+
if device and device.type == 'cuda':
|
72 |
+
torch.cuda.empty_cache()
|
73 |
+
gc.collect()
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
try:
|
76 |
print("Loading tokenizer...")
|
77 |
tokenizer = load_tokenizer()
|
78 |
+
if tokenizer:
|
79 |
+
print("Tokenizer loaded successfully!")
|
80 |
+
else:
|
81 |
+
print("Failed to load tokenizer!")
|
82 |
|
83 |
print("Loading model...")
|
84 |
model = load_model()
|
85 |
+
if model:
|
86 |
+
print("Model loaded successfully!")
|
87 |
+
else:
|
88 |
+
print("Failed to load model!")
|
89 |
|
90 |
except Exception as e:
|
91 |
print(f"Error during initialization: {str(e)}")
|
|
|
116 |
codes = data['codes']
|
117 |
if not isinstance(codes, list) or len(codes) == 0:
|
118 |
return jsonify({"error": "'codes' must be a non-empty array"}), 400
|
119 |
+
|
120 |
+
if len(codes) > 100:
|
121 |
+
return jsonify({"error": "Too many codes. Maximum 100 allowed."}), 400
|
122 |
+
|
123 |
+
validated_codes = []
|
124 |
+
for i, code in enumerate(codes):
|
125 |
+
if not isinstance(code, str):
|
126 |
+
return jsonify({"error": f"Code at index {i} must be a string"}), 400
|
127 |
+
if len(code.strip()) == 0:
|
128 |
+
validated_codes.append("# empty code")
|
129 |
+
elif len(code) > 50000:
|
130 |
+
return jsonify({"error": f"Code at index {i} too long. Maximum 50000 characters."}), 400
|
131 |
+
else:
|
132 |
+
validated_codes.append(code.strip())
|
133 |
+
|
134 |
+
if len(validated_codes) == 1:
|
135 |
+
score = predict_vulnerability_with_chunking(validated_codes[0])
|
136 |
+
cleanup_gpu_memory()
|
137 |
+
return jsonify({"results": [{"score": score}]})
|
138 |
+
|
139 |
+
batch_size = min(len(validated_codes), 16)
|
140 |
results = []
|
141 |
|
142 |
+
try:
|
143 |
+
for i in range(0, len(validated_codes), batch_size):
|
144 |
+
batch = validated_codes[i:i+batch_size]
|
145 |
+
|
146 |
+
long_codes = []
|
147 |
+
short_codes = []
|
148 |
+
long_indices = []
|
149 |
+
short_indices = []
|
150 |
+
|
151 |
+
for idx, code in enumerate(batch):
|
152 |
+
try:
|
153 |
+
tokens = tokenizer.encode(code, add_special_tokens=False, max_length=1000, truncation=True)
|
154 |
+
if len(tokens) > 450:
|
155 |
+
long_codes.append(code)
|
156 |
+
long_indices.append(i + idx)
|
157 |
+
else:
|
158 |
+
short_codes.append(code)
|
159 |
+
short_indices.append(i + idx)
|
160 |
+
except Exception as e:
|
161 |
+
print(f"Tokenization error for code {i + idx}: {e}")
|
162 |
+
short_codes.append(code)
|
163 |
+
short_indices.append(i + idx)
|
164 |
+
|
165 |
+
batch_scores = [0.0] * len(batch)
|
166 |
+
|
167 |
+
if short_codes:
|
168 |
+
try:
|
169 |
+
short_scores = predict_vulnerability_batch(short_codes)
|
170 |
+
for j, score in enumerate(short_scores):
|
171 |
+
local_idx = short_indices[j] - i
|
172 |
+
batch_scores[local_idx] = score
|
173 |
+
except Exception as e:
|
174 |
+
print(f"Batch prediction error: {e}")
|
175 |
+
for j in range(len(short_codes)):
|
176 |
+
local_idx = short_indices[j] - i
|
177 |
+
batch_scores[local_idx] = 0.0
|
178 |
+
|
179 |
+
for j, code in enumerate(long_codes):
|
180 |
+
try:
|
181 |
+
score = predict_vulnerability_with_chunking(code)
|
182 |
+
local_idx = long_indices[j] - i
|
183 |
+
batch_scores[local_idx] = score
|
184 |
+
except Exception as e:
|
185 |
+
print(f"Chunking prediction error: {e}")
|
186 |
+
local_idx = long_indices[j] - i
|
187 |
+
batch_scores[local_idx] = 0.0
|
188 |
+
|
189 |
+
for score in batch_scores:
|
190 |
+
results.append({"score": score})
|
191 |
+
|
192 |
+
cleanup_gpu_memory()
|
193 |
+
|
194 |
+
except Exception as e:
|
195 |
+
cleanup_gpu_memory()
|
196 |
+
raise e
|
197 |
|
198 |
return jsonify({"results": results})
|
199 |
|
200 |
except Exception as e:
|
201 |
+
cleanup_gpu_memory()
|
202 |
return jsonify({"error": f"Batch prediction error: {str(e)}"}), 500
|
203 |
|
204 |
+
def predict_vulnerability_with_chunking(code):
|
205 |
+
try:
|
206 |
+
if not code or len(code.strip()) == 0:
|
207 |
+
return 0.0
|
208 |
+
|
209 |
+
tokens = tokenizer.encode(code, add_special_tokens=False, max_length=2000, truncation=True)
|
210 |
+
|
211 |
+
if len(tokens) <= 450:
|
212 |
+
return predict_vulnerability(code)
|
213 |
+
|
214 |
+
chunk_size = 400
|
215 |
+
overlap = 50
|
216 |
+
max_score = 0.0
|
217 |
+
|
218 |
+
for start in range(0, len(tokens), chunk_size - overlap):
|
219 |
+
end = min(start + chunk_size, len(tokens))
|
220 |
+
chunk_tokens = tokens[start:end]
|
221 |
+
|
222 |
+
try:
|
223 |
+
chunk_code = tokenizer.decode(chunk_tokens, skip_special_tokens=True)
|
224 |
+
if chunk_code.strip():
|
225 |
+
score = predict_vulnerability(chunk_code)
|
226 |
+
max_score = max(max_score, score)
|
227 |
+
except Exception as e:
|
228 |
+
print(f"Chunk processing error: {e}")
|
229 |
+
continue
|
230 |
+
|
231 |
+
if end >= len(tokens):
|
232 |
+
break
|
233 |
+
|
234 |
+
return max_score
|
235 |
+
|
236 |
+
except Exception as e:
|
237 |
+
print(f"Chunking error: {e}")
|
238 |
+
return 0.0
|
239 |
|
240 |
def predict_vulnerability(code):
|
241 |
+
try:
|
242 |
+
if not code or len(code.strip()) == 0:
|
243 |
+
return 0.0
|
244 |
+
|
245 |
+
dynamic_length = min(max(len(code.split()) * 2, 128), 512)
|
246 |
+
|
247 |
+
inputs = tokenizer(
|
248 |
+
code,
|
249 |
+
truncation=True,
|
250 |
+
padding='max_length',
|
251 |
+
max_length=dynamic_length,
|
252 |
+
return_tensors='pt'
|
253 |
+
)
|
254 |
+
|
255 |
+
inputs = {k: v.to(device) for k, v in inputs.items()}
|
256 |
+
|
257 |
+
with torch.no_grad():
|
258 |
+
if device.type == 'cuda':
|
259 |
+
with torch.cuda.amp.autocast():
|
260 |
+
outputs = model(**inputs)
|
261 |
+
else:
|
262 |
+
outputs = model(**inputs)
|
263 |
+
|
264 |
+
if hasattr(outputs, 'logits'):
|
265 |
+
score = torch.sigmoid(outputs.logits).cpu().item()
|
266 |
+
else:
|
267 |
+
score = torch.sigmoid(outputs[0]).cpu().item()
|
268 |
+
|
269 |
+
return round(max(0.0, min(1.0, score)), 4)
|
270 |
+
|
271 |
+
except Exception as e:
|
272 |
+
print(f"Single prediction error: {e}")
|
273 |
+
return 0.0
|
274 |
|
275 |
def predict_vulnerability_batch(codes):
|
276 |
+
try:
|
277 |
+
if not codes or len(codes) == 0:
|
278 |
+
return []
|
279 |
+
|
280 |
+
filtered_codes = [code if code and code.strip() else "# empty" for code in codes]
|
281 |
+
|
282 |
+
max_len = max([len(code.split()) * 2 for code in filtered_codes if code])
|
283 |
+
dynamic_length = min(max(max_len, 128), 512)
|
284 |
+
|
285 |
+
inputs = tokenizer(
|
286 |
+
filtered_codes,
|
287 |
+
truncation=True,
|
288 |
+
padding='max_length',
|
289 |
+
max_length=dynamic_length,
|
290 |
+
return_tensors='pt'
|
291 |
+
)
|
292 |
+
|
293 |
+
inputs = {k: v.to(device) for k, v in inputs.items()}
|
294 |
+
|
295 |
+
with torch.no_grad():
|
296 |
+
if device.type == 'cuda':
|
297 |
+
with torch.cuda.amp.autocast():
|
298 |
+
outputs = model(**inputs)
|
299 |
+
else:
|
300 |
+
outputs = model(**inputs)
|
301 |
+
|
302 |
+
if hasattr(outputs, 'logits'):
|
303 |
+
scores = torch.sigmoid(outputs.logits).cpu().numpy()
|
304 |
+
else:
|
305 |
+
scores = torch.sigmoid(outputs[0]).cpu().numpy()
|
306 |
+
|
307 |
+
return [round(max(0.0, min(1.0, float(score))), 4) for score in scores.flatten()]
|
308 |
+
|
309 |
+
except Exception as e:
|
310 |
+
print(f"Batch prediction error: {e}")
|
311 |
+
return [0.0] * len(codes)
|
312 |
|
313 |
@app.route("/health", methods=['GET'])
|
314 |
def health_check():
|