lyi029 commited on
Commit
d4cb90d
·
1 Parent(s): ac3ca2a

performance

Browse files
Files changed (1) hide show
  1. app.py +115 -16
app.py CHANGED
@@ -15,8 +15,33 @@ import io
15
  import numpy as np
16
  from datasets import load_dataset
17
 
18
- # ===== 新增:SOM 相關 import =====
19
- from datasets import load_dataset
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
20
 
21
  # ⚠️ 注意: som_analyzer.py 必須存在於同一目錄
22
  try:
@@ -372,6 +397,13 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
372
 
373
  predictions = []
374
  all_embeddings = []
 
 
 
 
 
 
 
375
 
376
  try:
377
  with torch.no_grad():
@@ -396,7 +428,11 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
396
  inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
397
 
398
  # === Forward ===
399
- outputs = model(**inputs)
 
 
 
 
400
  logits = outputs.logits
401
 
402
  # === Softmax 機率 ===
@@ -414,11 +450,21 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
414
  print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
415
 
416
  # === Embedding ===
417
- if hasattr(outputs, "hidden_states") and outputs.hidden_states is not None:
418
- last_hidden = outputs.hidden_states[-1]
419
- cls_embedding = last_hidden[:, 0, :].cpu().numpy()[0]
420
- else:
421
- cls_embedding = logits.cpu().numpy()[0]
 
 
 
 
 
 
 
 
 
 
422
 
423
  # === Attention ===
424
  max_attention_score = 0.0
@@ -564,6 +610,8 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
564
  # 將 SOM 結果附加到最終回傳
565
  result["som_analysis"] = som_analysis
566
  print(f"✅ SOM analysis complete.")
 
 
567
 
568
  return result
569
 
@@ -605,16 +653,57 @@ def _split_into_basic_blocks(tokens):
605
  blocks.append(cur)
606
  return blocks
607
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
608
  def _block_attention_score(tokens, chunk_size=512):
609
  """
610
- 對單一 block 計算注意力:
611
- - 以 512 token 為單位分塊
612
- - 取最後一層 attention → head 平均 → 整體平均
613
- - 多個 chunk 取平均
 
614
  """
615
  scores = []
 
616
  for i in range(0, len(tokens), chunk_size):
617
  chunk = tokens[i:i+chunk_size]
 
618
  inputs = tokenizer(
619
  chunk,
620
  return_tensors="pt",
@@ -624,13 +713,23 @@ def _block_attention_score(tokens, chunk_size=512):
624
  padding=False
625
  )
626
  inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
627
- with torch.no_grad():
628
- out = model(**inputs, output_attentions=True)
629
- last_attn = out.attentions[-1] # (1, heads, seq, seq)
630
- mean_attn = last_attn.mean(dim=1).mean() # scalar
 
 
 
 
 
 
 
 
631
  scores.append(mean_attn.item())
 
632
  return float(np.mean(scores)) if scores else 0.0
633
 
 
634
  @app.post("/attention-heatmap")
635
  async def attention_heatmap(file: UploadFile = File(...)):
636
  """
 
15
  import numpy as np
16
  from datasets import load_dataset
17
 
18
+ import time
19
+ import psutil
20
+ global_heatmap_time = []
21
+
22
+ def measure(name: str):
23
+ """
24
+ 通用效能測量工具:
25
+ 用法:
26
+ with measure("inference"):
27
+ outputs = model(**inputs)
28
+ """
29
+ class PerfBlock:
30
+ def __enter__(self):
31
+ self.start = time.perf_counter()
32
+ self.process = psutil.Process()
33
+ self.mem_before = self.process.memory_info().rss
34
+ return self
35
+
36
+ def __exit__(self, exc_type, exc_val, exc_tb):
37
+ self.end = time.perf_counter()
38
+ self.duration_ms = (self.end - self.start) * 1000
39
+ self.mem_after = self.process.memory_info().rss
40
+ self.mem_usage = self.mem_after - self.mem_before
41
+ print(f"⏱️ [{name}] {self.duration_ms:.2f} ms, RAM +{self.mem_usage/1024:.1f} KB")
42
+
43
+ return PerfBlock()
44
+
45
 
46
  # ⚠️ 注意: som_analyzer.py 必須存在於同一目錄
47
  try:
 
397
 
398
  predictions = []
399
  all_embeddings = []
400
+
401
+ performance = {
402
+ "inference_ms": 0.0,
403
+ "embedding_ms": 0.0,
404
+ "heatmap_ms": 0.0
405
+ }
406
+
407
 
408
  try:
409
  with torch.no_grad():
 
428
  inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
429
 
430
  # === Forward ===
431
+ with measure("inference") as m:
432
+ outputs = model(**inputs)
433
+
434
+ performance["inference_ms"] += m.duration_ms
435
+ #outputs = model(**inputs)
436
  logits = outputs.logits
437
 
438
  # === Softmax 機率 ===
 
450
  print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
451
 
452
  # === Embedding ===
453
+ # if hasattr(outputs, "hidden_states") and outputs.hidden_states is not None:
454
+ # last_hidden = outputs.hidden_states[-1]
455
+ # cls_embedding = last_hidden[:, 0, :].cpu().numpy()[0]
456
+ # else:
457
+ # cls_embedding = logits.cpu().numpy()[0]
458
+ with measure("embedding") as m:
459
+ if hasattr(outputs, "hidden_states"):
460
+ last_hidden = outputs.hidden_states[-1]
461
+ cls_embedding = last_hidden[:, 0, :].cpu().numpy()[0]
462
+ else:
463
+ cls_embedding = logits.cpu().numpy()[0]
464
+
465
+ performance["embedding_ms"] += m.duration_ms
466
+
467
+
468
 
469
  # === Attention ===
470
  max_attention_score = 0.0
 
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
  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
  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
  """