FishServer/FishAction/PYTORCHVIDEO_TRAINING.md

PyTorchVideo Training (No SlowFast)

Your environment currently has a PyTorchVideo / torchvision mismatch that breaks pytorchvideo.transforms. So we train with PyTorchVideo decoding + a small custom transform pipeline (no SlowFast).

1) Prereqs

• You already created CSV files (e.g. train.csv, val.csv) with:
  • relative/path/to/video.mp4 <label_int>
• Your videos live under:
  • /home/ubuntu/data/fish/fish_action_videos
• Your CSV folder is:
  • /home/ubuntu/projects/FishAction/data/fish/fish_action_training_dataset

2) Train (fine-tune pretrained X3D)

From repo root:

cd /home/ubuntu/projects/FishAction

python train_pytorchvideo_x3d.py \
  --csv_dir /home/ubuntu/projects/FishAction/data/fish/fish_action_training_dataset \
  --path_prefix /home/ubuntu/data/fish/fish_action_videos \
  --model x3d_m \
  --pretrained \
  --num_frames 16 \
  --sampling_rate 5 \
  --batch_size 4 \
  --epochs 30 \
  --num_workers 4 \
  --amp \
  --output_dir /home/ubuntu/projects/FishAction/checkpoints/ptv_x3d_m

Notes:

• --pretrained uses torch.hub and will download weights to ~/.cache/torch/hub/ on first run.
• If you hit OOM, lower --batch_size to 2 or 1.

3) Outputs

The script writes into --output_dir:

• config.json
• checkpoint_last.pt
• checkpoint_best.pt