"""Batch runner: streaming multi-view weather transfer over EVERY sequence in a validation split. Loads the keyframe + interpolation pipes ONCE (via build_pipes) and loops run_sequence over all sequences, so the ~19 GB of weights + UMT5 text encoding are paid a single time instead of per sequence. Resumable: a sequence whose grid.mp4 already exists is skipped. Per-sequence failures are logged and skipped so one bad clip doesn't kill the whole run. Videos are expected pre-downloaded on disk as: //.mp4 (the seq id may contain '/', e.g. 'chunk_00.parquet/ip602_...'). Outputs go to //. Run inside the cosmos-predict2-local docker with A4D_WAN_FORCE_SDPA=1 and PYTHONPATH=/workspace (see custom/eval/vid2vid_eval/README.md). Example: A4D_WAN_FORCE_SDPA=1 PYTHONPATH=/workspace python custom/eval/vid2vid_eval/batch_val_transfer.py \ --kf-ckpt /assets/ckpts/kf.pt --inp-ckpt /assets/ckpts/inp.pt \ --wan-dir /assets/wan --vae-path /assets/wan/Wan2.1_VAE.pth \ --split-json /assets/split_val_transfer.json --data-root /assets/val_videos \ --cameras front_wide front_tele --num-frames 100 --out /assets/output/val_transfer """ import argparse import json import os import sys import time import traceback # repo root = .../Any4D ; this file lives at custom/eval/vid2vid_eval/ (4 levels down) ANY4D_ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))) if ANY4D_ROOT not in sys.path: sys.path.insert(0, ANY4D_ROOT) from custom.eval.vid2vid_eval.infer_multiview_video_transfer import ( build_pipes, run_sequence, SNOW_PROMPT, NEG_PROMPT) def _load_seqs(split_json, seqs_file): if seqs_file: return [l.strip() for l in open(seqs_file) if l.strip()] return list(json.load(open(split_json))['sequences'].keys()) def main(): p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) p.add_argument('--kf-ckpt', required=True) p.add_argument('--inp-ckpt', required=True) p.add_argument('--wan-dir', required=True) p.add_argument('--vae-path', default=None) p.add_argument('--kf-model', choices=['anchor', 'syncfz'], default='anchor', help='keyframe backend (anchor=cross-view Fun-Control, syncfz=pure Wan-T2V).') p.add_argument('--split-json', default=None, help='validation split json with a "sequences" dict.') p.add_argument('--seqs-file', default=None, help='newline-separated seq ids (overrides --split-json).') p.add_argument('--data-root', required=True, help='//.mp4') p.add_argument('--cameras', nargs='+', default=['front_wide', 'front_tele'], help='cameras per sequence; the FIRST is front_wide.') p.add_argument('--num-frames', type=int, default=100, help='10 s @ 10 fps = 100.') p.add_argument('--out', required=True, help='root output dir; each sequence -> //.') p.add_argument('--limit', type=int, default=0, help='process at most N sequences (0 = all).') p.add_argument('--overwrite', action='store_true', help='re-run even if /grid.mp4 exists.') # tuning passthroughs (defaults match the single-sequence script) p.add_argument('--prompt', default=SNOW_PROMPT) p.add_argument('--neg-prompt', default=NEG_PROMPT) p.add_argument('--sn-cutoff', type=float, default=20.0) p.add_argument('--sn-channels', type=int, default=4) p.add_argument('--kf-steps', type=int, default=35) p.add_argument('--kf-guidance', type=float, default=5.0) p.add_argument('--inp-steps', type=int, default=35) p.add_argument('--inp-guidance', type=float, default=5.0) p.add_argument('--seed', type=int, default=0) p.add_argument('--sdedit-strength', type=float, default=1.0) p.add_argument('--no-tone-match', action='store_true', help='disable per-keyframe tone-match (on by default).') p.add_argument('--front-w', type=int, default=1280) p.add_argument('--fps', type=int, default=10) args = p.parse_args() if not args.split_json and not args.seqs_file: raise SystemExit('provide --split-json or --seqs-file') seqs = _load_seqs(args.split_json, args.seqs_file) if args.limit: seqs = seqs[:args.limit] out_root = args.out if os.path.isabs(args.out) else os.path.join(ANY4D_ROOT, args.out) print(f'[batch] {len(seqs)} sequences, cameras={args.cameras}, num_frames={args.num_frames}') print(f'[batch] data-root={args.data_root} out={out_root}') pipes = build_pipes(args.kf_ckpt, args.inp_ckpt, args.wan_dir, args.vae_path, kf_model=args.kf_model, sn_cutoff=args.sn_cutoff, sn_channels=args.sn_channels, prompt=args.prompt, neg_prompt=args.neg_prompt) done, skipped, failed = 0, 0, 0 t0 = time.time() for si, seq in enumerate(seqs): out_dir = os.path.join(out_root, seq) grid = os.path.join(out_dir, 'grid.mp4') if not args.overwrite and os.path.exists(grid) and os.path.getsize(grid) > 0: print(f'[batch {si+1}/{len(seqs)}] SKIP (exists) {seq}') skipped += 1 continue videos = [os.path.join(args.data_root, seq, f'{c}.mp4') for c in args.cameras] missing = [v for v in videos if not os.path.isfile(v)] if missing: print(f'[batch {si+1}/{len(seqs)}] FAIL (missing videos) {seq}: {missing}') failed += 1 continue print(f'\n[batch {si+1}/{len(seqs)}] {seq} (done={done} skip={skipped} fail={failed}, ' f'elapsed={ (time.time()-t0)/60:.1f} min)') try: run_sequence(pipes, videos, args.num_frames, out_dir, kf_steps=args.kf_steps, kf_guidance=args.kf_guidance, inp_steps=args.inp_steps, inp_guidance=args.inp_guidance, seed=args.seed, sdedit_strength=args.sdedit_strength, tone_match=not args.no_tone_match, front_w=args.front_w, fps=args.fps) done += 1 except Exception: failed += 1 print(f'[batch {si+1}/{len(seqs)}] ERROR on {seq}:\n{traceback.format_exc()}') print(f'\n[batch DONE] done={done} skipped={skipped} failed={failed} ' f'total_time={(time.time()-t0)/60:.1f} min -> {out_root}') if __name__ == '__main__': main()