lyi029 commited on
Commit
11a7573
·
1 Parent(s): d4cb90d

performance

Browse files
Files changed (1) hide show
  1. app.py +13 -61
app.py CHANGED
@@ -17,7 +17,6 @@ from datasets import load_dataset
17
 
18
  import time
19
  import psutil
20
- global_heatmap_time = []
21
 
22
  def measure(name: str):
23
  """
@@ -610,8 +609,6 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
610
  # 將 SOM 結果附加到最終回傳
611
  result["som_analysis"] = som_analysis
612
  print(f"✅ SOM analysis complete.")
613
- result["performance"] = performance
614
-
615
 
616
  return result
617
 
@@ -653,57 +650,22 @@ def _split_into_basic_blocks(tokens):
653
  blocks.append(cur)
654
  return blocks
655
 
656
- # def _block_attention_score(tokens, chunk_size=512):
657
- # """
658
- # 對單一 block 計算注意力:
659
- # - 以 512 token 為單位分塊
660
- # - 取最後一層 attention → head 平均 → 整體平均
661
- # - 多個 chunk 取平均
662
- # """
663
- # scores = []
664
- # with measure("heatmap_block") as m:
665
- # out = model(**inputs, output_attentions=True)
666
-
667
- # # 可累積到總 heatmap 時間
668
- # global_heatmap_time.append(m.duration_ms)
669
-
670
- # for i in range(0, len(tokens), chunk_size):
671
- # chunk = tokens[i:i+chunk_size]
672
- # inputs = tokenizer(
673
- # chunk,
674
- # return_tensors="pt",
675
- # is_split_into_words=True,
676
- # truncation=True,
677
- # max_length=chunk_size,
678
- # padding=False
679
- # )
680
- # inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
681
- # with torch.no_grad():
682
- # out = model(**inputs, output_attentions=True)
683
- # last_attn = out.attentions[-1] # (1, heads, seq, seq)
684
- # mean_attn = last_attn.mean(dim=1).mean() # scalar
685
- # scores.append(mean_attn.item())
686
- # performance_heatmap = sum(global_heatmap_time)
687
- # global_heatmap_time.clear()
688
-
689
- # return float(np.mean(scores)) if scores else 0.0
690
-
691
-
692
- global_heatmap_time = []
693
-
694
  def _block_attention_score(tokens, chunk_size=512):
695
  """
696
- 計算單一 block 的 attention:
697
- - 512 tokens 一段
698
- - 計算 attention 平均值
699
- - 回傳整段 block 的 attention 平均
700
- 同時記錄 heatmap 計算的時間(ms)
701
  """
702
  scores = []
 
 
 
 
 
703
 
704
  for i in range(0, len(tokens), chunk_size):
705
  chunk = tokens[i:i+chunk_size]
706
-
707
  inputs = tokenizer(
708
  chunk,
709
  return_tensors="pt",
@@ -713,23 +675,13 @@ def _block_attention_score(tokens, chunk_size=512):
713
  padding=False
714
  )
715
  inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
716
-
717
- # === 計算 block attention(含效能測試) ===
718
- with measure("heatmap_block") as m:
719
- with torch.no_grad():
720
- out = model(**inputs, output_attentions=True)
721
-
722
- # 記錄每次 block 的時間
723
- global_heatmap_time.append(m.duration_ms)
724
-
725
- # attention score
726
- last_attn = out.attentions[-1] # (1, heads, seq, seq)
727
- mean_attn = last_attn.mean(dim=1).mean() # scalar
728
  scores.append(mean_attn.item())
729
-
730
  return float(np.mean(scores)) if scores else 0.0
731
 
732
-
733
  @app.post("/attention-heatmap")
734
  async def attention_heatmap(file: UploadFile = File(...)):
735
  """
 
17
 
18
  import time
19
  import psutil
 
20
 
21
  def measure(name: str):
22
  """
 
609
  # 將 SOM 結果附加到最終回傳
610
  result["som_analysis"] = som_analysis
611
  print(f"✅ SOM analysis complete.")
 
 
612
 
613
  return result
614
 
 
650
  blocks.append(cur)
651
  return blocks
652
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
653
  def _block_attention_score(tokens, chunk_size=512):
654
  """
655
+ 對單一 block 計算注意力:
656
+ - 以 512 token 為單位分塊
657
+ - 取最後一層 attention → head 平均 → 整體平均
658
+ - 多個 chunk 取平均
 
659
  """
660
  scores = []
661
+ with measure("heatmap_block") as m:
662
+ out = model(**inputs, output_attentions=True)
663
+
664
+ # 可累積到總 heatmap 時間
665
+ global_heatmap_time.append(m.duration_ms)
666
 
667
  for i in range(0, len(tokens), chunk_size):
668
  chunk = tokens[i:i+chunk_size]
 
669
  inputs = tokenizer(
670
  chunk,
671
  return_tensors="pt",
 
675
  padding=False
676
  )
677
  inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
678
+ with torch.no_grad():
679
+ out = model(**inputs, output_attentions=True)
680
+ last_attn = out.attentions[-1] # (1, heads, seq, seq)
681
+ mean_attn = last_attn.mean(dim=1).mean() # scalar
 
 
 
 
 
 
 
 
682
  scores.append(mean_attn.item())
 
683
  return float(np.mean(scores)) if scores else 0.0
684
 
 
685
  @app.post("/attention-heatmap")
686
  async def attention_heatmap(file: UploadFile = File(...)):
687
  """