FlashVTG/scripts/verify_tokenizer_alignment.py

#!/usr/bin/env python3
"""
验证 LLaMA tokenizer 与 qvhighlight_llama_text_feature 的对齐。
依据：extract_vidstg_llama_features.py、FlashVTG/start_end_dataset.py
"""
import os
import sys
import json

_project_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, _project_root)

def main():
    from transformers import AutoTokenizer
    import torch

    # 1. 与 extract_vidstg_llama_features.py 一致的 tokenizer
    # 代码引用: scripts/extract_vidstg_llama_features.py:36
    #   inp = tokenizer(text, return_tensors="pt", truncation=True, max_length=max_length)
    # 默认 add_special_tokens=True
    tok = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b", trust_remote_code=True)
    max_length = 40

    # 2. LLaMA special tokens (open_llama)
    print("=== LLaMA tokenizer special tokens ===")
    print(f"  bos_token: {repr(tok.bos_token)} (id={tok.bos_token_id})")
    print(f"  eos_token: {repr(tok.eos_token)} (id={tok.eos_token_id})")
    print(f"  pad_token: {repr(tok.pad_token)} (id={tok.pad_token_id})")

    # 3. 取 qvhighlight 中的样本
    vmr_path = os.path.join(_project_root, "data", "highlight_val_release.jsonl")
    feat_dir = os.path.join(_project_root, "features", "qvhighlight_llama_text_feature")
    if not os.path.exists(vmr_path):
        vmr_path = os.path.join(_project_root, "data", "highlight_val_release_IV2.jsonl")
    if not os.path.exists(vmr_path):
        print("No vmr file found, using hardcoded sample")
        samples = [{"qid": 2579, "query": "A girl and her mother cooked while talking with each other on facetime."}]
    else:
        with open(vmr_path) as f:
            samples = [json.loads(line) for line in f][:3]

    print("\n=== Tokenizer vs Feature alignment ===")
    print("  (attn 列数 nt = model L_txt，以 nt 为权威；viz 使用 target_len=nt)")
    for s in samples:
        qid, query = s["qid"], s["query"]
        # 与 extract_vidstg_llama_features 一致的编码
        enc = tok(query, truncation=True, max_length=max_length, add_special_tokens=True)
        input_ids = enc["input_ids"]
        tokens = tok.convert_ids_to_tokens(input_ids)
        tok_len = len(tokens)

        # 与 start_end_dataset._get_query_feat_by_qid 一致：加载 .pt
        # 代码引用: FlashVTG/start_end_dataset.py:477
        #   q_feat_path = join(self.q_feat_dir, f"qid{qid}.pt")
        #   q_feat = torch.load(q_feat_path).float().numpy()
        pt_path = os.path.join(feat_dir, f"qid{qid}.pt")
        if os.path.exists(pt_path):
            feat = torch.load(pt_path, map_location="cpu", weights_only=False)
            seq_len = feat.shape[0] if hasattr(feat, "shape") else len(feat)
            match = "OK" if tok_len == seq_len else "MISMATCH"
            print(f"  qid{qid}: tokenizer={tok_len}, .pt seq_len={seq_len} -> {match}")
            print(f"    first 3 tokens: {tokens[:3]}")
            print(f"    last 3 tokens:  {tokens[-3:]}")
            if tokens and tok.bos_token_id is not None:
                first_id = input_ids[0]
                is_bos = first_id == tok.bos_token_id
                print(f"    first token is BOS: {is_bos} (id={first_id})")
        else:
            print(f"  qid{qid}: .pt not found at {pt_path}")

if __name__ == "__main__":
    main()