InternVideo2_6B_V5_ACT75_eval.py

#!/usr/bin/env python
# coding: utf-8
# Cleaned and enhanced InternVideo2 6B evaluation script with structured logging
# Source:

import os
import sys
import subprocess
import logging
import json
import argparse
from pathlib import Path

import numpy as np
import cv2
import torch
from tqdm import tqdm
from huggingface_hub import hf_hub_download, HfApi, login


def setup_logging(log_level=logging.INFO, log_file=None):
    handlers = []
    fmt = '%(asctime)s - %(name)s - %(levelname)s - %(message)s'
    if log_file:
        handlers.append(logging.FileHandler(log_file))
    handlers.append(logging.StreamHandler(sys.stdout))
    logging.basicConfig(level=log_level, format=fmt, handlers=handlers)
    logging.info("Logging initialized.")


def run_command(cmd, cwd=None):
    logging.debug(f"Running command: {cmd} (cwd={cwd})")
    result = subprocess.run(cmd, shell=True, cwd=cwd, capture_output=True, text=True)
    if result.returncode != 0:
        logging.error(f"Command failed: {cmd}\nSTDOUT: {result.stdout}\nSTDERR: {result.stderr}")
        raise RuntimeError(f"Command '{cmd}' failed (exit code {result.returncode})")
    logging.debug(f"Command succeeded, output: {result.stdout.strip()}")
    return result.stdout.strip()


def download_checkpoint(repo_id: str, filename: str) -> str:
    logging.info(f"Downloading {filename} from {repo_id}...")
    path = hf_hub_download(repo_id=repo_id, filename=filename)
    logging.info(f"Downloaded vision checkpoint to {path}")
    return path


def load_config(config_path: str, vision_ckpt_path: str):
    from demo.config import Config, eval_dict_leaf
    logging.info(f"Loading config from {config_path}")
    cfg = Config.from_file(config_path)
    cfg = eval_dict_leaf(cfg)
    cfg.model.vision_ckpt_path = vision_ckpt_path
    cfg.model.vision_encoder.pretrained = vision_ckpt_path
    cfg.pretrained_path = vision_ckpt_path
    logging.debug(f"Config loaded: {cfg}")
    return cfg


def process_videos(
    json_path: str,
    model,
    config,
    output_prefix: str,
    num_frames_override: int = None
):
    """
    Run inference over each video, write outputs.
    If num_frames_override is given, use it; otherwise use config.num_frames.
    """
    from demo.utils import retrieve_text, _frame_from_video

    logging.info(f"Reading evaluation data from {json_path}")
    data = json.loads(Path(json_path).read_text())
    preds, logits = [], []

    # choose frame window size
    num_frames = num_frames_override if num_frames_override is not None else config.num_frames
    logging.info(f"Using window size: {num_frames} frames")

    for video_path, phrase, _ in data:
        logging.info("\n--- Starting new video ---")
        full_video = Path("photography-model") / video_path
        logging.info(f"Processing {full_video} with phrase '{phrase}'")
        frames = list(_frame_from_video(cv2.VideoCapture(str(full_video))))
        scores = []

        for j in tqdm(range(len(frames) - (num_frames - 1)), desc=Path(video_path).stem):
            _, probs = retrieve_text(
                frames[j : j + num_frames], [phrase],
                model=model, topk=1, config=config
            )
            scores.append(probs[0])

        best_idx = int(np.argmax(scores) + 1)
        preds.append(best_idx)
        logits.append(list(zip(map(float, scores), range(1, len(scores) + 1))))
        logging.info(f"Video result: predicted frame {best_idx}\n")

    preds_file = f"{output_prefix}-t{num_frames}.json"
    logits_file = f"{output_prefix}-logits-t{num_frames}.json"
    logging.info(f"Writing predictions to {preds_file}")
    Path(preds_file).write_text(json.dumps(preds, indent=2))
    logging.info(f"Writing logits to {logits_file}")
    Path(logits_file).write_text(json.dumps(logits, indent=2))

    return preds_file, logits_file


def upload_results(token: str, upload_files: list, repo_id: str):
    logging.info("Logging into Hugging Face Hub...")
    login(token)
    api = HfApi()
    for file_path in upload_files:
        logging.info(f"Uploading {file_path} to {repo_id}")
        api.upload_file(
            path_or_fileobj=file_path,
            path_in_repo=Path(file_path).name,
            repo_id=repo_id,
            repo_type="dataset",
        )
    logging.info("Upload complete.")


def main():
    parser = argparse.ArgumentParser(
        description="Evaluate InternVideo2 sliding-window retrieval."
    )
    parser.add_argument(
        "--branch",
        type=str,
        default="main",
        help="Branch to use for evaluation."
    )
    parser.add_argument(
        "--num_frames",
        type=int,
        default=None,
        help="Manually set the number of frames per window."
    )
    args = parser.parse_args()

    setup_logging()

    # ensure IV2 repo
    iv2_path = Path('~/IV2').expanduser()
    if not iv2_path.exists():
        logging.info("Cloning IV2 repository...")
        run_command('git clone https://github.com/qingy1337/IV2.git ~/IV2')

    os.chdir(iv2_path / 'InternVideo2' / 'multi_modality')
    sys.path.append(os.getcwd())
    run_command(f'git checkout {args.branch}', cwd=os.getcwd())

    MODEL_NAME = '6B'
    vision_ckpt = download_checkpoint(
        repo_id="OpenGVLab/InternVideo2-Stage2_6B-224p-f4",
        filename="internvideo2-s2_6b-224p-f4.pt"
    )
    config = load_config('scripts/pretraining/stage2/6B/config.py', vision_ckpt)
    from demo.utils import setup_internvideo2
    model, tokenizer = setup_internvideo2(config)

    if not Path('photography-model').exists():
        run_command('git clone https://github.com/ruo2019/photography-model.git')

    prefix = f"ACT75-V5-InternVideo-{MODEL_NAME}"
    preds_file, logits_file = process_videos(
        'photography-model/data/ACT75.json',
        model, config, prefix,
        num_frames_override=args.num_frames
    )

    upload_results(
        os.getenv('HF_TOKEN', ''),
        [preds_file, logits_file],
        'qingy2024/InternVideo2-Data'
    )


if __name__ == '__main__':
    main()