updates
Browse files
app.py
CHANGED
|
@@ -386,14 +386,72 @@ def _block_attention_score(tokens, chunk_size=512):
|
|
| 386 |
scores.append(mean_attn.item())
|
| 387 |
return float(np.mean(scores)) if scores else 0.0
|
| 388 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 389 |
@app.post("/attention-heatmap")
|
| 390 |
async def attention_heatmap(file: UploadFile = File(...)):
|
| 391 |
"""
|
| 392 |
適用格式:整行以 '</s>' 分隔的 disasm tokens。
|
| 393 |
每個 block 以 jump/call/ret 結束。
|
| 394 |
最多取前 1000 個 block。
|
|
|
|
|
|
|
|
|
|
|
|
|
| 395 |
"""
|
| 396 |
-
import io, sys, numpy as np, csv
|
| 397 |
csv.field_size_limit(sys.maxsize)
|
| 398 |
|
| 399 |
try:
|
|
@@ -429,16 +487,42 @@ async def attention_heatmap(file: UploadFile = File(...)):
|
|
| 429 |
"score": score
|
| 430 |
})
|
| 431 |
|
| 432 |
-
|
| 433 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 434 |
return {
|
| 435 |
"num_blocks": len(results),
|
| 436 |
"avg_attention": avg_score,
|
| 437 |
-
"blocks": results
|
|
|
|
|
|
|
| 438 |
}
|
| 439 |
|
| 440 |
except Exception as e:
|
| 441 |
print(f"❌ attention_heatmap error: {e}")
|
|
|
|
|
|
|
| 442 |
return {"error": str(e)}
|
| 443 |
|
| 444 |
|
|
|
|
| 386 |
scores.append(mean_attn.item())
|
| 387 |
return float(np.mean(scores)) if scores else 0.0
|
| 388 |
|
| 389 |
+
# @app.post("/attention-heatmap")
|
| 390 |
+
# async def attention_heatmap(file: UploadFile = File(...)):
|
| 391 |
+
# """
|
| 392 |
+
# 適用格式:整行以 '</s>' 分隔的 disasm tokens。
|
| 393 |
+
# 每個 block 以 jump/call/ret 結束。
|
| 394 |
+
# 最多取前 1000 個 block。
|
| 395 |
+
# """
|
| 396 |
+
# import io, sys, numpy as np, csv
|
| 397 |
+
# csv.field_size_limit(sys.maxsize)
|
| 398 |
+
|
| 399 |
+
# try:
|
| 400 |
+
# # === 讀檔 ===
|
| 401 |
+
# content = await file.read()
|
| 402 |
+
# text = content.decode("utf-8", errors="ignore")
|
| 403 |
+
|
| 404 |
+
# # === 以 </s> 分隔整個 disasm ===
|
| 405 |
+
# lines = text.replace("\n", " ").split("</s>")
|
| 406 |
+
# tokens = []
|
| 407 |
+
# for line in lines:
|
| 408 |
+
# toks = [t.strip() for t in line.split() if t.strip()]
|
| 409 |
+
# if toks:
|
| 410 |
+
# tokens.extend(toks)
|
| 411 |
+
|
| 412 |
+
# print(f"📦 Total tokens parsed: {len(tokens)}")
|
| 413 |
+
|
| 414 |
+
# # === 拆 basic blocks ===
|
| 415 |
+
# blocks = _split_into_basic_blocks(tokens)
|
| 416 |
+
# print(f"📊 Total blocks before truncation: {len(blocks)}")
|
| 417 |
+
|
| 418 |
+
# # === 只取前 1000 個 block
|
| 419 |
+
# MAX_BLOCKS = int(os.getenv("MAX_HEATMAP_BLOCKS", "1000"))
|
| 420 |
+
# blocks = blocks[:MAX_BLOCKS]
|
| 421 |
+
|
| 422 |
+
# # === 計算每個 block 的 attention
|
| 423 |
+
# results = []
|
| 424 |
+
# for i, blk in enumerate(blocks):
|
| 425 |
+
# score = _block_attention_score(blk)
|
| 426 |
+
# results.append({
|
| 427 |
+
# "idx": i,
|
| 428 |
+
# "size": len(blk),
|
| 429 |
+
# "score": score
|
| 430 |
+
# })
|
| 431 |
+
|
| 432 |
+
# avg_score = float(np.mean([b["score"] for b in results])) if results else 0.0
|
| 433 |
+
|
| 434 |
+
# return {
|
| 435 |
+
# "num_blocks": len(results),
|
| 436 |
+
# "avg_attention": avg_score,
|
| 437 |
+
# "blocks": results
|
| 438 |
+
# }
|
| 439 |
+
|
| 440 |
+
# except Exception as e:
|
| 441 |
+
# print(f"❌ attention_heatmap error: {e}")
|
| 442 |
+
# return {"error": str(e)}
|
| 443 |
@app.post("/attention-heatmap")
|
| 444 |
async def attention_heatmap(file: UploadFile = File(...)):
|
| 445 |
"""
|
| 446 |
適用格式:整行以 '</s>' 分隔的 disasm tokens。
|
| 447 |
每個 block 以 jump/call/ret 結束。
|
| 448 |
最多取前 1000 個 block。
|
| 449 |
+
會同時回傳:
|
| 450 |
+
- blocks:每個 basic block 的注意力平均
|
| 451 |
+
- attention_heatmap:base64 圖像
|
| 452 |
+
- heatmap_matrix:數值陣列(可供前端繪圖)
|
| 453 |
"""
|
| 454 |
+
import io, sys, numpy as np, csv, base64, matplotlib.pyplot as plt
|
| 455 |
csv.field_size_limit(sys.maxsize)
|
| 456 |
|
| 457 |
try:
|
|
|
|
| 487 |
"score": score
|
| 488 |
})
|
| 489 |
|
| 490 |
+
if not results:
|
| 491 |
+
return {"error": "No valid blocks parsed"}
|
| 492 |
+
|
| 493 |
+
scores = [b["score"] for b in results]
|
| 494 |
+
avg_score = float(np.mean(scores))
|
| 495 |
+
|
| 496 |
+
# === 視覺化成 heatmap ===
|
| 497 |
+
fig, ax = plt.subplots(figsize=(10, 2))
|
| 498 |
+
heatmap_data = np.expand_dims(scores, axis=0)
|
| 499 |
+
im = ax.imshow(heatmap_data, cmap="viridis", aspect="auto")
|
| 500 |
+
plt.colorbar(im, ax=ax, orientation="vertical", label="Attention Score")
|
| 501 |
+
ax.set_title("Attention Heatmap per Basic Block")
|
| 502 |
+
ax.set_xlabel("Block Index")
|
| 503 |
+
ax.set_yticks([]) # 不需要 Y 軸
|
| 504 |
+
plt.tight_layout()
|
| 505 |
+
|
| 506 |
+
# === 轉成 base64 ===
|
| 507 |
+
buf = io.BytesIO()
|
| 508 |
+
plt.savefig(buf, format="png")
|
| 509 |
+
buf.seek(0)
|
| 510 |
+
img_base64 = base64.b64encode(buf.read()).decode("utf-8")
|
| 511 |
+
plt.close(fig)
|
| 512 |
+
|
| 513 |
+
# === 回傳 JSON ===
|
| 514 |
return {
|
| 515 |
"num_blocks": len(results),
|
| 516 |
"avg_attention": avg_score,
|
| 517 |
+
"blocks": results,
|
| 518 |
+
"heatmap_matrix": heatmap_data.tolist(), # ✅ 可互動繪圖
|
| 519 |
+
"attention_heatmap": f"data:image/png;base64,{img_base64}" # ✅ 前端直接顯示
|
| 520 |
}
|
| 521 |
|
| 522 |
except Exception as e:
|
| 523 |
print(f"❌ attention_heatmap error: {e}")
|
| 524 |
+
import traceback
|
| 525 |
+
traceback.print_exc()
|
| 526 |
return {"error": str(e)}
|
| 527 |
|
| 528 |
|