111306047 commited on
Commit
8dcc174
·
1 Parent(s): 6ff5685
Files changed (1) hide show
  1. app.py +62 -0
app.py CHANGED
@@ -74,6 +74,67 @@ def get_model_info():
74
  "id2label_sample": id2label_sample # ✅ 除錯用
75
  }
76
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77
  @app.post("/predict")
78
  async def predict_malware(files: List[UploadFile] = File(...)):
79
  """
@@ -205,6 +266,7 @@ async def predict_malware(files: List[UploadFile] = File(...)):
205
 
206
  print(f"🎯 Selected embedding from: {best_embedding['filename']}")
207
  print(f" Attention score: {best_embedding['attention_score']:.4f}")
 
208
 
209
  # ===== 返回結果 =====
210
  result = {
 
74
  "id2label_sample": id2label_sample # ✅ 除錯用
75
  }
76
 
77
+ # ===== 載入預先降維的資料 =====
78
+ # 在模型載入時執行一次
79
+ print("📊 Loading pre-computed t-SNE embeddings...")
80
+ df_tsne = pd.read_json("tsne_points.json") # 你上傳的 JSON
81
+ tsne_coords = df_tsne[['x', 'y']].values # (N, 2)
82
+
83
+ # 載入原始 embeddings(用於 kNN 投影)
84
+ df_original = pd.read_csv("cobert_finetune_embedding_checkpoint-126000.csv")
85
+ dim_cols = [f'emb_{i}' for i in range(768)]
86
+ original_embeddings = df_original[dim_cols].to_numpy() # (N, 768)
87
+
88
+ # 標準化(必須和訓練時一致)
89
+ scaler = StandardScaler()
90
+ original_embeddings_scaled = scaler.fit_transform(original_embeddings)
91
+
92
+ # 建立 kNN 模型(用於投影新點)
93
+ knn_model = NearestNeighbors(n_neighbors=10, metric='euclidean')
94
+ knn_model.fit(original_embeddings_scaled)
95
+
96
+ print(f"✅ Loaded {len(df_tsne)} pre-computed t-SNE points")
97
+
98
+
99
+ def project_new_embedding_to_tsne(new_embedding_768d, k=10):
100
+ """
101
+ 將新的 768 維 embedding 投影到已有的 t-SNE 空間
102
+
103
+ 方法:找到 k 個最近鄰,用距離加權平均它們的 t-SNE 座標
104
+
105
+ Args:
106
+ new_embedding_768d: array of shape (768,)
107
+ k: 使用幾個最近鄰
108
+
109
+ Returns:
110
+ dict: {'x': float, 'y': float}
111
+ """
112
+ # 1. 標準化新 embedding(使用相同的 scaler)
113
+ new_emb_scaled = scaler.transform([new_embedding_768d]) # (1, 768)
114
+
115
+ # 2. 找到 k 個最近鄰
116
+ distances, indices = knn_model.kneighbors(new_emb_scaled, n_neighbors=k)
117
+ distances = distances[0] # (k,)
118
+ indices = indices[0] # (k,)
119
+
120
+ # 3. 距離加權平均(距離越近權重越大)
121
+ epsilon = 1e-8
122
+ weights = 1 / (distances + epsilon)
123
+ weights = weights / weights.sum()
124
+
125
+ # 4. 計算加權平均的 t-SNE 座標
126
+ neighbor_coords = tsne_coords[indices] # (k, 2)
127
+ projected_x = np.dot(weights, neighbor_coords[:, 0])
128
+ projected_y = np.dot(weights, neighbor_coords[:, 1])
129
+
130
+ return {
131
+ 'x': float(projected_x),
132
+ 'y': float(projected_y),
133
+ 'method': 'kNN_projection',
134
+ 'k': k,
135
+ 'nearest_distance': float(distances[0])
136
+ }
137
+
138
  @app.post("/predict")
139
  async def predict_malware(files: List[UploadFile] = File(...)):
140
  """
 
266
 
267
  print(f"🎯 Selected embedding from: {best_embedding['filename']}")
268
  print(f" Attention score: {best_embedding['attention_score']:.4f}")
269
+
270
 
271
  # ===== 返回結果 =====
272
  result = {