updates
Browse files
app.py
CHANGED
|
@@ -30,7 +30,7 @@ app.add_middleware(
|
|
| 30 |
MODEL_REPO = os.getenv("MODEL_REPO", "lyi029/codebert-finetuned")
|
| 31 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 32 |
DEVICE = torch.device("cpu")
|
| 33 |
-
HF_DATASET_REPO = "lyi029/malvec-embeddings"
|
| 34 |
|
| 35 |
# ====== 連接 Space B ======
|
| 36 |
SPACE_B_URL = os.getenv("SPACE_B_URL", "https://lyi029-model-update.hf.space")
|
|
@@ -425,6 +425,193 @@ def upload_embeddings_to_hub(local_path: str):
|
|
| 425 |
|
| 426 |
# except Exception as e:
|
| 427 |
# raise HTTPException(status_code=500, detail=f"Prediction failed: {e}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 428 |
@app.post("/predict")
|
| 429 |
async def predict_malware(files: List[UploadFile] = File(...)):
|
| 430 |
"""
|
|
@@ -447,9 +634,8 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 447 |
try:
|
| 448 |
content = await file.read()
|
| 449 |
asm_code = content.decode('utf-8').strip()
|
| 450 |
-
|
| 451 |
if not asm_code:
|
| 452 |
-
print(f"⚠️
|
| 453 |
continue
|
| 454 |
|
| 455 |
print(f"📄 [{idx}/{len(files)}] Processing: {file.filename}")
|
|
@@ -467,16 +653,19 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 467 |
# === Forward ===
|
| 468 |
outputs = model(**inputs)
|
| 469 |
logits = outputs.logits
|
|
|
|
|
|
|
| 470 |
probs = torch.nn.functional.softmax(logits, dim=-1)
|
| 471 |
-
|
| 472 |
|
|
|
|
|
|
|
| 473 |
predicted_id = predicted_class.item()
|
| 474 |
label_name = (
|
| 475 |
model.config.id2label[predicted_id]
|
| 476 |
if hasattr(model.config, "id2label") and predicted_id in model.config.id2label
|
| 477 |
else f"class_{predicted_id}"
|
| 478 |
)
|
| 479 |
-
|
| 480 |
print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
|
| 481 |
|
| 482 |
# === Embedding ===
|
|
@@ -503,11 +692,12 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 503 |
})
|
| 504 |
|
| 505 |
all_embeddings.append({
|
| 506 |
-
"
|
| 507 |
"label": label_name,
|
| 508 |
"embedding": cls_embedding.tolist(),
|
| 509 |
"attention_score": max_attention_score,
|
| 510 |
-
"text_excerpt": asm_code
|
|
|
|
| 511 |
})
|
| 512 |
|
| 513 |
print(f" ✅ Predicted: {label_name} (confidence: {confidence.item():.3f})")
|
|
@@ -522,20 +712,20 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 522 |
# === 多數投票 ===
|
| 523 |
label_votes = [p["predicted_label"] for p in predictions]
|
| 524 |
vote_counts = Counter(label_votes)
|
| 525 |
-
final_label = vote_counts.most_common(1)[0]
|
| 526 |
-
final_count = vote_counts[final_label]
|
| 527 |
|
| 528 |
print(f"\n📊 Vote distribution: {dict(vote_counts)}")
|
| 529 |
print(f"🏆 Final prediction: {final_label} ({final_count}/{len(predictions)} votes)")
|
| 530 |
|
| 531 |
-
# === 最高注意力權重的 embedding ===
|
| 532 |
same_label_embeddings = [emb for emb in all_embeddings if emb["label"] == final_label]
|
| 533 |
-
|
| 534 |
-
|
| 535 |
-
|
| 536 |
-
|
|
|
|
| 537 |
|
| 538 |
-
print(f"🎯 Selected embedding from: {best_embedding['
|
| 539 |
print(f" Attention score: {best_embedding['attention_score']:.4f}")
|
| 540 |
|
| 541 |
# === 回傳結果 ===
|
|
@@ -547,7 +737,7 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 547 |
"vote_distribution": dict(vote_counts),
|
| 548 |
"embedding": {
|
| 549 |
"values": best_embedding["embedding"],
|
| 550 |
-
"source_file": best_embedding["
|
| 551 |
"attention_score": best_embedding["attention_score"],
|
| 552 |
"dimension": len(best_embedding["embedding"])
|
| 553 |
},
|
|
@@ -557,7 +747,6 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 557 |
print(f"\n✅ Prediction complete!\n")
|
| 558 |
|
| 559 |
# === 寫入本地 embeddings_log.jsonl ===
|
| 560 |
-
os.makedirs("results", exist_ok=True)
|
| 561 |
log_path = os.path.join("results", "embeddings_log.jsonl")
|
| 562 |
with open(log_path, "a") as f:
|
| 563 |
for emb in all_embeddings:
|
|
@@ -565,7 +754,7 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 565 |
print(f"📝 Logged {len(all_embeddings)} embeddings locally.")
|
| 566 |
|
| 567 |
# === 上傳到 Hugging Face Dataset ===
|
| 568 |
-
repo_id = "lyi029/malvec-embeddings"
|
| 569 |
api = HfApi()
|
| 570 |
|
| 571 |
try:
|
|
@@ -582,11 +771,11 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 582 |
except Exception:
|
| 583 |
print("ℹ️ No previous embeddings found in dataset.")
|
| 584 |
|
| 585 |
-
# 合併去重
|
| 586 |
with open(log_path, "r") as f:
|
| 587 |
new_data = [json.loads(l) for l in f if l.strip()]
|
| 588 |
|
| 589 |
-
|
|
|
|
| 590 |
merged = list(merged_dict.values())
|
| 591 |
|
| 592 |
with open(log_path, "w") as f:
|
|
@@ -612,6 +801,7 @@ async def predict_malware(files: List[UploadFile] = File(...)):
|
|
| 612 |
traceback.print_exc()
|
| 613 |
raise HTTPException(status_code=500, detail=f"Prediction failed: {str(e)}")
|
| 614 |
|
|
|
|
| 615 |
# ----------------------------------------------------------
|
| 616 |
# 手動觸發 Space B 訓練
|
| 617 |
# ----------------------------------------------------------
|
|
|
|
| 30 |
MODEL_REPO = os.getenv("MODEL_REPO", "lyi029/codebert-finetuned")
|
| 31 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 32 |
DEVICE = torch.device("cpu")
|
| 33 |
+
HF_DATASET_REPO = "lyi029/malvec-embeddings-new"
|
| 34 |
|
| 35 |
# ====== 連接 Space B ======
|
| 36 |
SPACE_B_URL = os.getenv("SPACE_B_URL", "https://lyi029-model-update.hf.space")
|
|
|
|
| 425 |
|
| 426 |
# except Exception as e:
|
| 427 |
# raise HTTPException(status_code=500, detail=f"Prediction failed: {e}")
|
| 428 |
+
# @app.post("/predict")
|
| 429 |
+
# async def predict_malware(files: List[UploadFile] = File(...)):
|
| 430 |
+
# """
|
| 431 |
+
# 批次預測惡意軟體家族(維持原回傳格式)
|
| 432 |
+
# 並自動寫入 embeddings_log.jsonl、上傳到 Hugging Face Dataset
|
| 433 |
+
# """
|
| 434 |
+
# if not files:
|
| 435 |
+
# raise HTTPException(status_code=400, detail="No files provided")
|
| 436 |
+
|
| 437 |
+
# print(f"\n{'='*60}")
|
| 438 |
+
# print(f"📥 Received {len(files)} files")
|
| 439 |
+
# print(f"{'='*60}")
|
| 440 |
+
|
| 441 |
+
# predictions = []
|
| 442 |
+
# all_embeddings = []
|
| 443 |
+
|
| 444 |
+
# try:
|
| 445 |
+
# with torch.no_grad():
|
| 446 |
+
# for idx, file in enumerate(files, 1):
|
| 447 |
+
# try:
|
| 448 |
+
# content = await file.read()
|
| 449 |
+
# asm_code = content.decode('utf-8').strip()
|
| 450 |
+
|
| 451 |
+
# if not asm_code:
|
| 452 |
+
# print(f"⚠️ Skip empty file: {file.filename}")
|
| 453 |
+
# continue
|
| 454 |
+
|
| 455 |
+
# print(f"📄 [{idx}/{len(files)}] Processing: {file.filename}")
|
| 456 |
+
|
| 457 |
+
# # === Tokenize ===
|
| 458 |
+
# inputs = tokenizer(
|
| 459 |
+
# asm_code,
|
| 460 |
+
# return_tensors="pt",
|
| 461 |
+
# max_length=512,
|
| 462 |
+
# truncation=True,
|
| 463 |
+
# padding=True
|
| 464 |
+
# )
|
| 465 |
+
# inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
|
| 466 |
+
|
| 467 |
+
# # === Forward ===
|
| 468 |
+
# outputs = model(**inputs)
|
| 469 |
+
# logits = outputs.logits
|
| 470 |
+
# probs = torch.nn.functional.softmax(logits, dim=-1)
|
| 471 |
+
# confidence, predicted_class = torch.max(probs, dim=-1)
|
| 472 |
+
|
| 473 |
+
# predicted_id = predicted_class.item()
|
| 474 |
+
# label_name = (
|
| 475 |
+
# model.config.id2label[predicted_id]
|
| 476 |
+
# if hasattr(model.config, "id2label") and predicted_id in model.config.id2label
|
| 477 |
+
# else f"class_{predicted_id}"
|
| 478 |
+
# )
|
| 479 |
+
|
| 480 |
+
# print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
|
| 481 |
+
|
| 482 |
+
# # === Embedding ===
|
| 483 |
+
# if hasattr(outputs, "hidden_states") and outputs.hidden_states is not None:
|
| 484 |
+
# last_hidden = outputs.hidden_states[-1]
|
| 485 |
+
# cls_embedding = last_hidden[:, 0, :].cpu().numpy()[0]
|
| 486 |
+
# else:
|
| 487 |
+
# cls_embedding = logits.cpu().numpy()[0]
|
| 488 |
+
|
| 489 |
+
# # === Attention ===
|
| 490 |
+
# max_attention_score = 0.0
|
| 491 |
+
# if hasattr(outputs, "attentions") and outputs.attentions is not None:
|
| 492 |
+
# last_attention = outputs.attentions[-1]
|
| 493 |
+
# avg_attention = last_attention.mean(dim=1)
|
| 494 |
+
# cls_attention = avg_attention[0, 0, :].cpu().numpy()
|
| 495 |
+
# max_attention_score = float(cls_attention.max())
|
| 496 |
+
|
| 497 |
+
# # === Collect Results ===
|
| 498 |
+
# predictions.append({
|
| 499 |
+
# "filename": file.filename,
|
| 500 |
+
# "predicted_label": label_name,
|
| 501 |
+
# "confidence": float(confidence.item()),
|
| 502 |
+
# "attention_score": max_attention_score
|
| 503 |
+
# })
|
| 504 |
+
|
| 505 |
+
# all_embeddings.append({
|
| 506 |
+
# "source": file.filename,
|
| 507 |
+
# "label": label_name,
|
| 508 |
+
# "embedding": cls_embedding.tolist(),
|
| 509 |
+
# "attention_score": max_attention_score,
|
| 510 |
+
# "text_excerpt": asm_code # 方便後續分析
|
| 511 |
+
# })
|
| 512 |
+
|
| 513 |
+
# print(f" ✅ Predicted: {label_name} (confidence: {confidence.item():.3f})")
|
| 514 |
+
|
| 515 |
+
# except Exception as e:
|
| 516 |
+
# print(f" ❌ Error processing {file.filename}: {e}")
|
| 517 |
+
# continue
|
| 518 |
+
|
| 519 |
+
# if not predictions:
|
| 520 |
+
# raise HTTPException(status_code=400, detail="No valid files processed")
|
| 521 |
+
|
| 522 |
+
# # === 多數投票 ===
|
| 523 |
+
# label_votes = [p["predicted_label"] for p in predictions]
|
| 524 |
+
# vote_counts = Counter(label_votes)
|
| 525 |
+
# final_label = vote_counts.most_common(1)[0][0]
|
| 526 |
+
# final_count = vote_counts[final_label]
|
| 527 |
+
|
| 528 |
+
# print(f"\n📊 Vote distribution: {dict(vote_counts)}")
|
| 529 |
+
# print(f"🏆 Final prediction: {final_label} ({final_count}/{len(predictions)} votes)")
|
| 530 |
+
|
| 531 |
+
# # === 最高注意力權重的 embedding ===
|
| 532 |
+
# same_label_embeddings = [emb for emb in all_embeddings if emb["label"] == final_label]
|
| 533 |
+
# if same_label_embeddings:
|
| 534 |
+
# best_embedding = max(same_label_embeddings, key=lambda x: x["attention_score"])
|
| 535 |
+
# else:
|
| 536 |
+
# best_embedding = max(all_embeddings, key=lambda x: x["attention_score"])
|
| 537 |
+
|
| 538 |
+
# print(f"🎯 Selected embedding from: {best_embedding['filename']}")
|
| 539 |
+
# print(f" Attention score: {best_embedding['attention_score']:.4f}")
|
| 540 |
+
|
| 541 |
+
# # === 回傳結果 ===
|
| 542 |
+
# result = {
|
| 543 |
+
# "final_label": final_label,
|
| 544 |
+
# "vote_count": final_count,
|
| 545 |
+
# "total_segments": len(predictions),
|
| 546 |
+
# "confidence": sum(p["confidence"] for p in predictions) / len(predictions),
|
| 547 |
+
# "vote_distribution": dict(vote_counts),
|
| 548 |
+
# "embedding": {
|
| 549 |
+
# "values": best_embedding["embedding"],
|
| 550 |
+
# "source_file": best_embedding["filename"],
|
| 551 |
+
# "attention_score": best_embedding["attention_score"],
|
| 552 |
+
# "dimension": len(best_embedding["embedding"])
|
| 553 |
+
# },
|
| 554 |
+
# "segment_predictions": predictions
|
| 555 |
+
# }
|
| 556 |
+
|
| 557 |
+
# print(f"\n✅ Prediction complete!\n")
|
| 558 |
+
|
| 559 |
+
# # === 寫入本地 embeddings_log.jsonl ===
|
| 560 |
+
# os.makedirs("results", exist_ok=True)
|
| 561 |
+
# log_path = os.path.join("results", "embeddings_log.jsonl")
|
| 562 |
+
# with open(log_path, "a") as f:
|
| 563 |
+
# for emb in all_embeddings:
|
| 564 |
+
# f.write(json.dumps(emb) + "\n")
|
| 565 |
+
# print(f"📝 Logged {len(all_embeddings)} embeddings locally.")
|
| 566 |
+
|
| 567 |
+
# # === 上傳到 Hugging Face Dataset ===
|
| 568 |
+
# repo_id = "lyi029/malvec-embeddings"
|
| 569 |
+
# api = HfApi()
|
| 570 |
+
|
| 571 |
+
# try:
|
| 572 |
+
# old_data = []
|
| 573 |
+
# try:
|
| 574 |
+
# old_path = hf_hub_download(
|
| 575 |
+
# repo_id=repo_id,
|
| 576 |
+
# filename="embeddings_log.jsonl",
|
| 577 |
+
# repo_type="dataset",
|
| 578 |
+
# token=HF_TOKEN
|
| 579 |
+
# )
|
| 580 |
+
# with open(old_path, "r") as f:
|
| 581 |
+
# old_data = [json.loads(l) for l in f if l.strip()]
|
| 582 |
+
# except Exception:
|
| 583 |
+
# print("ℹ️ No previous embeddings found in dataset.")
|
| 584 |
+
|
| 585 |
+
# # 合併去重
|
| 586 |
+
# with open(log_path, "r") as f:
|
| 587 |
+
# new_data = [json.loads(l) for l in f if l.strip()]
|
| 588 |
+
|
| 589 |
+
# merged_dict = {d["filename"]: d for d in (old_data + new_data)}
|
| 590 |
+
# merged = list(merged_dict.values())
|
| 591 |
+
|
| 592 |
+
# with open(log_path, "w") as f:
|
| 593 |
+
# for item in merged:
|
| 594 |
+
# f.write(json.dumps(item) + "\n")
|
| 595 |
+
|
| 596 |
+
# api.upload_file(
|
| 597 |
+
# path_or_fileobj=log_path,
|
| 598 |
+
# path_in_repo="embeddings_log.jsonl",
|
| 599 |
+
# repo_id=repo_id,
|
| 600 |
+
# repo_type="dataset",
|
| 601 |
+
# token=HF_TOKEN,
|
| 602 |
+
# )
|
| 603 |
+
# print(f"☁️ Uploaded {len(merged)} merged embeddings to dataset: {repo_id}")
|
| 604 |
+
# except Exception as e:
|
| 605 |
+
# print(f"⚠️ Dataset upload failed: {e}")
|
| 606 |
+
|
| 607 |
+
# return result
|
| 608 |
+
|
| 609 |
+
# except Exception as e:
|
| 610 |
+
# print(f"\n❌ Error in predict_malware: {e}")
|
| 611 |
+
# import traceback
|
| 612 |
+
# traceback.print_exc()
|
| 613 |
+
# raise HTTPException(status_code=500, detail=f"Prediction failed: {str(e)}")
|
| 614 |
+
|
| 615 |
@app.post("/predict")
|
| 616 |
async def predict_malware(files: List[UploadFile] = File(...)):
|
| 617 |
"""
|
|
|
|
| 634 |
try:
|
| 635 |
content = await file.read()
|
| 636 |
asm_code = content.decode('utf-8').strip()
|
|
|
|
| 637 |
if not asm_code:
|
| 638 |
+
print(f"⚠️ Skip empty file: {file.filename}")
|
| 639 |
continue
|
| 640 |
|
| 641 |
print(f"📄 [{idx}/{len(files)}] Processing: {file.filename}")
|
|
|
|
| 653 |
# === Forward ===
|
| 654 |
outputs = model(**inputs)
|
| 655 |
logits = outputs.logits
|
| 656 |
+
|
| 657 |
+
# === Softmax 機率 ===
|
| 658 |
probs = torch.nn.functional.softmax(logits, dim=-1)
|
| 659 |
+
prob_list = probs.squeeze().cpu().tolist()
|
| 660 |
|
| 661 |
+
# === 取得預測結果 ===
|
| 662 |
+
confidence, predicted_class = torch.max(probs, dim=-1)
|
| 663 |
predicted_id = predicted_class.item()
|
| 664 |
label_name = (
|
| 665 |
model.config.id2label[predicted_id]
|
| 666 |
if hasattr(model.config, "id2label") and predicted_id in model.config.id2label
|
| 667 |
else f"class_{predicted_id}"
|
| 668 |
)
|
|
|
|
| 669 |
print(f" 🔍 Predicted ID: {predicted_id}, Label: {label_name}")
|
| 670 |
|
| 671 |
# === Embedding ===
|
|
|
|
| 692 |
})
|
| 693 |
|
| 694 |
all_embeddings.append({
|
| 695 |
+
"source": file.filename,
|
| 696 |
"label": label_name,
|
| 697 |
"embedding": cls_embedding.tolist(),
|
| 698 |
"attention_score": max_attention_score,
|
| 699 |
+
"text_excerpt": asm_code,
|
| 700 |
+
"probs": prob_list # ✅ 新增:完整 softmax 機率
|
| 701 |
})
|
| 702 |
|
| 703 |
print(f" ✅ Predicted: {label_name} (confidence: {confidence.item():.3f})")
|
|
|
|
| 712 |
# === 多數投票 ===
|
| 713 |
label_votes = [p["predicted_label"] for p in predictions]
|
| 714 |
vote_counts = Counter(label_votes)
|
| 715 |
+
final_label, final_count = vote_counts.most_common(1)[0]
|
|
|
|
| 716 |
|
| 717 |
print(f"\n📊 Vote distribution: {dict(vote_counts)}")
|
| 718 |
print(f"🏆 Final prediction: {final_label} ({final_count}/{len(predictions)} votes)")
|
| 719 |
|
| 720 |
+
# === 選最高注意力權重的 embedding ===
|
| 721 |
same_label_embeddings = [emb for emb in all_embeddings if emb["label"] == final_label]
|
| 722 |
+
best_embedding = (
|
| 723 |
+
max(same_label_embeddings, key=lambda x: x["attention_score"])
|
| 724 |
+
if same_label_embeddings else
|
| 725 |
+
max(all_embeddings, key=lambda x: x["attention_score"])
|
| 726 |
+
)
|
| 727 |
|
| 728 |
+
print(f"🎯 Selected embedding from: {best_embedding['source']}")
|
| 729 |
print(f" Attention score: {best_embedding['attention_score']:.4f}")
|
| 730 |
|
| 731 |
# === 回傳結果 ===
|
|
|
|
| 737 |
"vote_distribution": dict(vote_counts),
|
| 738 |
"embedding": {
|
| 739 |
"values": best_embedding["embedding"],
|
| 740 |
+
"source_file": best_embedding["source"],
|
| 741 |
"attention_score": best_embedding["attention_score"],
|
| 742 |
"dimension": len(best_embedding["embedding"])
|
| 743 |
},
|
|
|
|
| 747 |
print(f"\n✅ Prediction complete!\n")
|
| 748 |
|
| 749 |
# === 寫入本地 embeddings_log.jsonl ===
|
|
|
|
| 750 |
log_path = os.path.join("results", "embeddings_log.jsonl")
|
| 751 |
with open(log_path, "a") as f:
|
| 752 |
for emb in all_embeddings:
|
|
|
|
| 754 |
print(f"📝 Logged {len(all_embeddings)} embeddings locally.")
|
| 755 |
|
| 756 |
# === 上傳到 Hugging Face Dataset ===
|
| 757 |
+
repo_id = "lyi029/malvec-embeddings-new" # ✅ 你的新 dataset
|
| 758 |
api = HfApi()
|
| 759 |
|
| 760 |
try:
|
|
|
|
| 771 |
except Exception:
|
| 772 |
print("ℹ️ No previous embeddings found in dataset.")
|
| 773 |
|
|
|
|
| 774 |
with open(log_path, "r") as f:
|
| 775 |
new_data = [json.loads(l) for l in f if l.strip()]
|
| 776 |
|
| 777 |
+
# 合併去重
|
| 778 |
+
merged_dict = {(d["source"]): d for d in (old_data + new_data)}
|
| 779 |
merged = list(merged_dict.values())
|
| 780 |
|
| 781 |
with open(log_path, "w") as f:
|
|
|
|
| 801 |
traceback.print_exc()
|
| 802 |
raise HTTPException(status_code=500, detail=f"Prediction failed: {str(e)}")
|
| 803 |
|
| 804 |
+
|
| 805 |
# ----------------------------------------------------------
|
| 806 |
# 手動觸發 Space B 訓練
|
| 807 |
# ----------------------------------------------------------
|