svd_train.yaml

environment:
  mode: 0
  debug: False
  seed: 42
  distributed: False
  enable_modelarts: False

train:
  epochs: 1000
  temporal_only: True   # train only the temporal layers or the entire UNet
  pretrained: ./model/svd-d19a808f.ckpt
  output_dir: output/
  save_interval: 10

  dataset:
    class_path: data.video_dataset.VideoDataset
    init_args:
      data_dir:
      metadata:
      frames: 4
      step: 3

  dataloader:
    batch_size: 1
    shuffle: True
    drop_remainder: True

  scheduler:
    name: cosine_decay
    lr: 3.0e-5
    end_lr: 1.0e-7
    warmup_steps: 50

  optimizer:
    name: adamw
    betas: [ 0.9, 0.98 ]
    weight_decay: 0.01

  settings:
    drop_overflow_update: True
    gradient_accumulation_steps: 5
    clip_grad: True
    clip_norm: 1.0

LossScale:
  loss_scale_value: 65536
  scale_factor: 2
  scale_window: 1000