update som
Browse files
app.py
CHANGED
|
@@ -269,36 +269,60 @@ def analyze_som_region_probability(
|
|
| 269 |
|
| 270 |
|
| 271 |
async def analyze_with_som(embedding, predicted_family, filename):
|
| 272 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 273 |
if not SOM_AVAILABLE:
|
| 274 |
print("⚠️ SOM analysis disabled")
|
| 275 |
return None
|
| 276 |
try:
|
| 277 |
-
print("\n🗺️ Performing SOM analysis...")
|
| 278 |
-
if isinstance(embedding, list):
|
| 279 |
-
embedding_array = np.array(embedding, dtype=np.float32)
|
| 280 |
-
else:
|
| 281 |
-
embedding_array = embedding
|
| 282 |
-
|
| 283 |
-
winner_row, winner_col = som_analyzer.find_winner_neuron(embedding_array)
|
| 284 |
-
print(f" 📍 Winner neuron: ({winner_row}, {winner_col})")
|
| 285 |
|
|
|
|
| 286 |
hf_embeddings = await get_embeddings_from_hf_dataset(filename)
|
| 287 |
|
|
|
|
|
|
|
|
|
|
| 288 |
if hf_embeddings and len(hf_embeddings) > 0:
|
| 289 |
-
|
| 290 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 291 |
som_analyzer, winner_row, winner_col, hf_embeddings, radius=2
|
| 292 |
)
|
| 293 |
-
data_source = "hf_dataset"
|
| 294 |
else:
|
| 295 |
-
|
|
|
|
| 296 |
features_dict = som_analyzer.predict_features_from_position(
|
| 297 |
winner_row, winner_col, radius=1
|
| 298 |
)
|
| 299 |
features_prob = {k: float(v) for k, v in features_dict.items()}
|
| 300 |
-
|
| 301 |
-
|
|
|
|
|
|
|
| 302 |
features_binary = {
|
| 303 |
feature: 1 if prob > 0.2 else 0
|
| 304 |
for feature, prob in features_prob.items()
|
|
@@ -306,7 +330,6 @@ async def analyze_with_som(embedding, predicted_family, filename):
|
|
| 306 |
print(f" ✅ Features detected: {features_binary}")
|
| 307 |
|
| 308 |
result = {
|
| 309 |
-
# 🐛 (from app.py) Bug Fix: col_winner -> winner_col
|
| 310 |
"winner_position": {"row": int(winner_row), "col": int(winner_col)},
|
| 311 |
"features": features_binary,
|
| 312 |
"features_probability": features_prob,
|
|
@@ -314,8 +337,11 @@ async def analyze_with_som(embedding, predicted_family, filename):
|
|
| 314 |
}
|
| 315 |
print(f"✅ SOM analysis complete!")
|
| 316 |
return result
|
|
|
|
| 317 |
except Exception as e:
|
| 318 |
print(f"❌ SOM analysis error: {e}")
|
|
|
|
|
|
|
| 319 |
return None
|
| 320 |
|
| 321 |
|
|
|
|
| 269 |
|
| 270 |
|
| 271 |
async def analyze_with_som(embedding, predicted_family, filename):
|
| 272 |
+
"""
|
| 273 |
+
(from app.py)
|
| 274 |
+
✨ 修改後的 SOM 分析邏輯:
|
| 275 |
+
1. 優先從 HF Dataset 抓取 embeddings。
|
| 276 |
+
2. 如果找到,用它們的「平均值」來決定座標。
|
| 277 |
+
3. 如果沒找到,才用模型即時產生的 embedding。
|
| 278 |
+
"""
|
| 279 |
if not SOM_AVAILABLE:
|
| 280 |
print("⚠️ SOM analysis disabled")
|
| 281 |
return None
|
| 282 |
try:
|
| 283 |
+
print("\n🗺️ Performing SOM analysis (New Logic)...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 284 |
|
| 285 |
+
# 1. 先去抓資料
|
| 286 |
hf_embeddings = await get_embeddings_from_hf_dataset(filename)
|
| 287 |
|
| 288 |
+
embedding_to_use = None
|
| 289 |
+
data_source = ""
|
| 290 |
+
|
| 291 |
if hf_embeddings and len(hf_embeddings) > 0:
|
| 292 |
+
# 2. 如果找到了 -> 計算平均 embedding
|
| 293 |
+
print(f"✅ Found {len(hf_embeddings)} embeddings from HF dataset. Calculating average.")
|
| 294 |
+
embedding_to_use = np.mean(hf_embeddings, axis=0, dtype=np.float32)
|
| 295 |
+
data_source = "hf_dataset_avg"
|
| 296 |
+
else:
|
| 297 |
+
# 3. Fallback: 使用模型即時產生的 embedding
|
| 298 |
+
print(f"⚠️ No HF data found, falling back to new embedding.")
|
| 299 |
+
if isinstance(embedding, list):
|
| 300 |
+
embedding_to_use = np.array(embedding, dtype=np.float32)
|
| 301 |
+
else:
|
| 302 |
+
embedding_to_use = embedding
|
| 303 |
+
data_source = "realtime_model" # (之前這裡是 background_data,現在更正)
|
| 304 |
+
|
| 305 |
+
# 4. 根據選定的 embedding 決定座標
|
| 306 |
+
winner_row, winner_col = som_analyzer.find_winner_neuron(embedding_to_use)
|
| 307 |
+
print(f" 📍 Winner neuron: ({winner_row}, {winner_col}) (Source: {data_source})")
|
| 308 |
+
|
| 309 |
+
# 5. 分析座標周圍 (使用已抓取的 hf_embeddings 或 fallback)
|
| 310 |
+
if hf_embeddings and len(hf_embeddings) > 0:
|
| 311 |
+
# (使用原有的鄰域分析邏輯)
|
| 312 |
+
features_prob = analyze_som_region_probability(
|
| 313 |
som_analyzer, winner_row, winner_col, hf_embeddings, radius=2
|
| 314 |
)
|
|
|
|
| 315 |
else:
|
| 316 |
+
# (使用原有的 fallback 邏輯)
|
| 317 |
+
print(f"⚠️ No HF data for prob, using background data")
|
| 318 |
features_dict = som_analyzer.predict_features_from_position(
|
| 319 |
winner_row, winner_col, radius=1
|
| 320 |
)
|
| 321 |
features_prob = {k: float(v) for k, v in features_dict.items()}
|
| 322 |
+
# 確保 data_source 反映這一點
|
| 323 |
+
data_source = "background_data_prob" if data_source == "realtime_model" else data_source
|
| 324 |
+
|
| 325 |
+
# 6. 組合最終結果 (此部分不變)
|
| 326 |
features_binary = {
|
| 327 |
feature: 1 if prob > 0.2 else 0
|
| 328 |
for feature, prob in features_prob.items()
|
|
|
|
| 330 |
print(f" ✅ Features detected: {features_binary}")
|
| 331 |
|
| 332 |
result = {
|
|
|
|
| 333 |
"winner_position": {"row": int(winner_row), "col": int(winner_col)},
|
| 334 |
"features": features_binary,
|
| 335 |
"features_probability": features_prob,
|
|
|
|
| 337 |
}
|
| 338 |
print(f"✅ SOM analysis complete!")
|
| 339 |
return result
|
| 340 |
+
|
| 341 |
except Exception as e:
|
| 342 |
print(f"❌ SOM analysis error: {e}")
|
| 343 |
+
import traceback
|
| 344 |
+
traceback.print_exc()
|
| 345 |
return None
|
| 346 |
|
| 347 |
|