configs/experiment/projects/bridging/dinosaur/coco_feat_rec_dino_base16_auto.yaml

# @package _global_
# ViT feature reconstruction on COCO using an autoregressive decoder (SLATE style)
defaults:
  - /experiment/projects/bridging/dinosaur/_base_feature_recon  # (1)!
  - /dataset: coco  # (2)!
  - /experiment/projects/bridging/dinosaur/_preprocessing_coco_dino_feature_recon_ccrop # (3)!
  - /experiment/projects/bridging/dinosaur/_metrics_coco # (4)!
  - _self_

# The following parameters assume training on 8 GPUs, leading to an effective batch size of 64.
trainer:
  devices: 8
  max_steps: 500000
  max_epochs:
  gradient_clip_val: 1.0

dataset:
  num_workers: 4
  batch_size: 8

models:
  conditioning:
    _target_: routed.ocl.conditioning.RandomConditioning
    n_slots: 7
    object_dim: 256

    batch_size_path: input.batch_size
  feature_extractor:
    model_name: vit_base_patch16_224_dino
    pretrained: ${when_testing:false,true}
    freeze: true

  perceptual_grouping: {}
  object_decoder:
    _target_: routed.ocl.decoding.AutoregressivePatchDecoder
    decoder_cond_dim: ${.output_dim}
    use_input_transform: true
    use_decoder_masks: true
    decoder:
      _target_: ocl.neural_networks.build_transformer_decoder
      _partial_: true
      n_layers: 4
      n_heads: 8
      return_attention_weights: true
    masks_path: perceptual_grouping.feature_attributions
    object_features_path: perceptual_grouping.objects
experiment:
  input_feature_dim: 768