configs/experiment/slot_attention/movi_c_optical_flow.yaml

# @package _global_
# Configuration to exactly reproduce unsupervised object recognition of the original slot attention
# paper.
defaults:
  - /experiment/slot_attention/_base_optical_flow  # (1)!
  - /dataset: movi_c_image  # (2)!
  - /experiment/slot_attention/_preprocessing_movi # (3)!
  - _self_

trainer:
  devices: 8
  max_steps: 500000
  max_epochs:

dataset:
  num_workers: 4
  batch_size: 8
  # Additionally sample frames from flow_backward.
  train_transforms:
    02_sample_frames:
      fields: [image, backward_flow]
    03_preprocessing:
      transforms:
        # Addtionally preprocess backward_flow.
        # TODO(flwenzel): Normalize flow?
        backward_flow:
          _target_: torchvision.transforms.Compose
          transforms:
            - _target_: ocl.preprocessing.FlowToTensor
            - _target_: torchvision.transforms.Resize
              size: 128
  eval_transforms:
    02_sample_frames:
      fields: [image, mask, backward_flow]
    03_preprocessing:
      transforms:
        # Addtionally preprocess backward_flow.
        backward_flow:
          _target_: torchvision.transforms.Compose
          transforms:
            - _target_: ocl.preprocessing.FlowToTensor
            - _target_: torchvision.transforms.Resize
              size: 128


models:
  conditioning:
    n_slots: 7