dry_run_train.py

import argparse
import json
import os
import sys
from types import SimpleNamespace


def ensure_dummy_dataset(base_dir: str, num_entries: int = 20) -> str:
    try:
        import numpy as np
        import imageio.v2 as imageio
    except Exception as exc:
        raise RuntimeError("imageio and numpy are required to build the dummy dataset.") from exc

    os.makedirs(base_dir, exist_ok=True)
    video_path = os.path.join(base_dir, "dummy.mp4")
    json_path = os.path.join(base_dir, "dummy_dataset.json")

    if not os.path.exists(video_path):
        writer = imageio.get_writer(video_path, fps=16)
        for i in range(90):
            frame = np.zeros((128, 128, 3), dtype=np.uint8)
            frame[:, :, 0] = (i * 3) % 255
            frame[:, :, 1] = (i * 7) % 255
            frame[:, :, 2] = (i * 13) % 255
            writer.append_data(frame)
        writer.close()

    facedetect_v1 = [
        [{"angle": {"yaw": 0, "pitch": 0, "roll": 0}, "detect": {"top": 10, "height": 100, "width": 100, "left": 10, "prob": 1.0}}],
        [{"angle": {"yaw": 60, "pitch": 0, "roll": 0}, "detect": {"top": 10, "height": 100, "width": 100, "left": 10, "prob": 1.0}}],
        [{"angle": {"yaw": -60, "pitch": 0, "roll": 0}, "detect": {"top": 10, "height": 100, "width": 100, "left": 10, "prob": 1.0}}],
    ]
    facedetect_v1_frame_index = [0, 20, 40]

    context = {
        "disk_path": video_path,
        "text": "dummy caption",
        "facedetect_v1": facedetect_v1,
        "facedetect_v1_frame_index": facedetect_v1_frame_index,
    }
    meta = {f"video_{i}": context for i in range(num_entries)}

    with open(json_path, "w") as f:
        json.dump(meta, f)

    return json_path


def build_args() -> SimpleNamespace:
    return SimpleNamespace(
        zero_face_ratio=0.0,
        shot_rope=False,
        split_rope=False,
        split1=False,
        split2=False,
        split3=False,
    )


def main() -> int:
    os.environ.setdefault("CUDA_VISIBLE_DEVICES", "0")
    parser = argparse.ArgumentParser(description="Dry-run the full training pipeline (single forward).")
    parser.add_argument("--json_path", type=str, default="", help="Existing dataset JSON path.")
    parser.add_argument("--resolution", type=int, nargs=2, default=[64, 64], help="Resolution as H W.")
    parser.add_argument("--ref_num", type=int, default=3, help="Reference image count.")
    parser.add_argument("--base_dir", type=str, default="/data/rczhang/PencilFolder/multi-shot/tmp_dryrun")
    parser.add_argument("--model_root", type=str, default="/data/rczhang/PencilFolder/DiffSynth-Studio/models")
    parser.add_argument("--model_id", type=str, default="Wan-AI/Wan2.2-TI2V-5B")
    args = parser.parse_args()

    if args.json_path:
        json_path = args.json_path
    else:
        json_path = ensure_dummy_dataset(args.base_dir)

    sys.path.insert(0, "/data/rczhang/PencilFolder/multi-shot")
    from multi_view.datasets.videodataset import MulltiShot_MultiView_Dataset
    from multi_view.train import WanTrainingModule

    dataset = MulltiShot_MultiView_Dataset(
        dataset_base_path=json_path,
        resolution=tuple(args.resolution),
        ref_num=args.ref_num,
        training=True,
    )
    batch = [dataset[0]]

    model_id_with_origin_paths = ",".join([
        f"{args.model_id}:diffusion_pytorch_model*.safetensors",
        f"{args.model_id}:models_t5_umt5-xxl-enc-bf16.pth",
        f"{args.model_id}:Wan2.2_VAE.pth",
    ])

    model = WanTrainingModule(
        model_id_with_origin_paths=model_id_with_origin_paths,
        trainable_models="dit",
        local_model_path=args.model_root,
    )

    import torch
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device)
    model.pipe.device = device
    model.pipe.torch_dtype = torch.bfloat16
    print("Dry-run device:", device)

    runtime_args = build_args()
    inputs = model.forward_preprocess(batch)
    loss = model.forward(batch, runtime_args, inputs=inputs)
    print("Dry-run loss:", float(loss.detach().cpu().item()))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())