configs/experiment/slate/movi_e.yaml

# @package _global_
# Configuration to exactly reproduce unsupervised object recognition of the original SLATE
# paper.
defaults:
  - /experiment/slate/_base  # (1)!
  - /dataset: movi_e_image  # (2)!
  - /experiment/slate/_movi_preprocessing # (3)!
  - /experiment/slate/_metrics_clevr_patch # (4)!
  - _self_

trainer:
  devices: 8
  max_steps: 200000
  max_epochs:

dataset:
  num_workers: 8
  batch_size: 8

models:
  conditioning:
    n_slots: 24

  object_decoder:
    decoder:
      n_layers: 8
      n_heads: 4
    masks_path: perceptual_grouping.feature_attributions

  perceptual_grouping:
    iters: 7

  masks_as_image:
    _target_: routed.ocl.utils.resizing.Resize
    input_path: object_decoder.masks
    size: 128
    resize_mode: bilinear
    patch_mode: true