python -m scripts.build_manifest --manifest runs/exp001/dataset_manifest.jsonl --roots H:/Videos/Cartoons --backend opencv --errors runs/exp001/manifest_errors.jsonl

Train:

python -m src.train --config config.yaml

Resume:

python -m src.train --config config.yaml --resume runs/exp001/checkpoints/last_full.pt

Inference:

python -m src.infer --config config.yaml --weights runs/exp001/checkpoints/last_weights.pt --input input.mp4 --output output_x2.mp4

Features

AMP: bf16 / fp16 / no
AdamW + cosine or constant scheduler + warmup
EMA weights for inference checkpoint
JSONL training logs
Full checkpoint for resume and separate weights-only checkpoint
Recursive dataset scanning + manifest caching

Notes

train_crop controls patch size. Increase for better quality if VRAM allows.
frame_stride_min/max control motion magnitude for training samples.
pad_multiple should be a multiple of 8 (or your attention window size).
performance.* toggles PyTorch SDPA backends (flash / mem-efficient / math).
data.backend selects the video backend; decord is usually fastest for random access.

Optional backends:

pip install decord
pip install av

Optional WebDataset:

pip install webdataset

Dataset preparation (SSD subset)

Create a smaller, subset on SSD with per-root limits:

python -m scripts.prepare_dataset --out path ^
  --root-limit "" ^
  --root-limit "" ^
  --root-limit "" ^
  --root-limit "" ^
  --mode copy --skip-existing --seed 42

Optional: write a manifest for the prepared subset:

python -m scripts.prepare_dataset --out path --root-limit path --write-manifest path/manifest.jsonl --manifest-backend decord

Triplet cache (NPZ)

Create a cache of pre-sampled triplets (much faster training, lower CPU usage):

python -m scripts.cache_triplets --config config.yaml --out path --count 200000 --batch-size 8 --num-workers 4 --backend decord

Then train from NPZ:

data:
  format: "npz"
  npz_dir: "E:/FrameNN_triplets"

Triplet cache (WebDataset)

Create WebDataset shards (PNG triplets inside .tar files):

python -m scripts.cache_wds --config config.yaml --out path --count 200000 --batch-size 4 --num-workers 3 --backend decord --shard-size 1000

Train from WebDataset:

data:
  format: "wds"
  wds_urls: "E:/FrameNN_wds/shard-{000000..000199}.tar"
  wds_shuffle: 1000

Model diagnostics

python -m src.diagnose --config config.yaml --weights path/to/model --input "" --max-pairs 80 --match-brightness