performance
Browse files
app.py
CHANGED
|
@@ -15,33 +15,6 @@ import io
|
|
| 15 |
import numpy as np
|
| 16 |
from datasets import load_dataset
|
| 17 |
|
| 18 |
-
import time
|
| 19 |
-
import psutil
|
| 20 |
-
|
| 21 |
-
def measure(name: str):
|
| 22 |
-
"""
|
| 23 |
-
通用效能測量工具:
|
| 24 |
-
用法:
|
| 25 |
-
with measure("inference"):
|
| 26 |
-
outputs = model(**inputs)
|
| 27 |
-
"""
|
| 28 |
-
class PerfBlock:
|
| 29 |
-
def __enter__(self):
|
| 30 |
-
self.start = time.perf_counter()
|
| 31 |
-
self.process = psutil.Process()
|
| 32 |
-
self.mem_before = self.process.memory_info().rss
|
| 33 |
-
return self
|
| 34 |
-
|
| 35 |
-
def __exit__(self, exc_type, exc_val, exc_tb):
|
| 36 |
-
self.end = time.perf_counter()
|
| 37 |
-
self.duration_ms = (self.end - self.start) * 1000
|
| 38 |
-
self.mem_after = self.process.memory_info().rss
|
| 39 |
-
self.mem_usage = self.mem_after - self.mem_before
|
| 40 |
-
print(f"⏱️ [{name}] {self.duration_ms:.2f} ms, RAM +{self.mem_usage/1024:.1f} KB")
|
| 41 |
-
|
| 42 |
-
return PerfBlock()
|
| 43 |
-
|
| 44 |
-
|
| 45 |
# ⚠️ 注意: som_analyzer.py 必須存在於同一目錄
|
| 46 |
try:
|
| 47 |
from som_analyzer import SOMAnalyzer
|
|
@@ -396,13 +369,6 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
|
|
| 396 |
|
| 397 |
predictions = []
|
| 398 |
all_embeddings = []
|
| 399 |
-
|
| 400 |
-
performance = {
|
| 401 |
-
"inference_ms": 0.0,
|
| 402 |
-
"embedding_ms": 0.0,
|
| 403 |
-
"heatmap_ms": 0.0
|
| 404 |
-
}
|
| 405 |
-
|
| 406 |
|
| 407 |
try:
|
| 408 |
with torch.no_grad():
|
|
@@ -427,11 +393,7 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
|
|
| 427 |
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
|
| 428 |
|
| 429 |
# === Forward ===
|
| 430 |
-
|
| 431 |
-
outputs = model(**inputs)
|
| 432 |
-
|
| 433 |
-
performance["inference_ms"] += m.duration_ms
|
| 434 |
-
#outputs = model(**inputs)
|
| 435 |
logits = outputs.logits
|
| 436 |
|
| 437 |
# === Softmax 機率 ===
|
|
@@ -449,21 +411,11 @@ async def predict_malware(files: List[UploadFile] = File(...), original_filename
|
|
| 449 |
print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
|
| 450 |
|
| 451 |
# === Embedding ===
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
|
| 456 |
-
|
| 457 |
-
with measure("embedding") as m:
|
| 458 |
-
if hasattr(outputs, "hidden_states"):
|
| 459 |
-
last_hidden = outputs.hidden_states[-1]
|
| 460 |
-
cls_embedding = last_hidden[:, 0, :].cpu().numpy()[0]
|
| 461 |
-
else:
|
| 462 |
-
cls_embedding = logits.cpu().numpy()[0]
|
| 463 |
-
|
| 464 |
-
performance["embedding_ms"] += m.duration_ms
|
| 465 |
-
|
| 466 |
-
|
| 467 |
|
| 468 |
# === Attention ===
|
| 469 |
max_attention_score = 0.0
|
|
@@ -658,12 +610,6 @@ def _block_attention_score(tokens, chunk_size=512):
|
|
| 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(
|
|
|
|
| 15 |
import numpy as np
|
| 16 |
from datasets import load_dataset
|
| 17 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
# ⚠️ 注意: som_analyzer.py 必須存在於同一目錄
|
| 19 |
try:
|
| 20 |
from som_analyzer import SOMAnalyzer
|
|
|
|
| 369 |
|
| 370 |
predictions = []
|
| 371 |
all_embeddings = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 372 |
|
| 373 |
try:
|
| 374 |
with torch.no_grad():
|
|
|
|
| 393 |
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
|
| 394 |
|
| 395 |
# === Forward ===
|
| 396 |
+
outputs = model(**inputs)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 397 |
logits = outputs.logits
|
| 398 |
|
| 399 |
# === Softmax 機率 ===
|
|
|
|
| 411 |
print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
|
| 412 |
|
| 413 |
# === Embedding ===
|
| 414 |
+
if hasattr(outputs, "hidden_states") and outputs.hidden_states is not None:
|
| 415 |
+
last_hidden = outputs.hidden_states[-1]
|
| 416 |
+
cls_embedding = last_hidden[:, 0, :].cpu().numpy()[0]
|
| 417 |
+
else:
|
| 418 |
+
cls_embedding = logits.cpu().numpy()[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 419 |
|
| 420 |
# === Attention ===
|
| 421 |
max_attention_score = 0.0
|
|
|
|
| 610 |
- 多個 chunk 取平均
|
| 611 |
"""
|
| 612 |
scores = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 613 |
for i in range(0, len(tokens), chunk_size):
|
| 614 |
chunk = tokens[i:i+chunk_size]
|
| 615 |
inputs = tokenizer(
|