eval_downstream_prediction.yaml

checkpoint_path: ???
device: cuda
seed: 12345
batch_size: 64
steps: 6000   # maximum number of steps
downstream_model: MLP1   # "linear" or "MLPx" or "wideMLPx" where x is a positive integer
learning_rate: 1e-3
lr_decay_milestones: [2000, 4000]
lr_decay_factor: 0.5

train_size: 10000
validation_size: 1024
test_size: 2000

validation_every: 250

# Matching can be: 'loss', 'mask' (only for object-centric), 'deterministic' (only for VAEs).
matching: loss

train_ignore_mode: default
test_ignore_mode: default

# Default dataset variant types to consider for downstream prediction task.
# If some of these are missing for the datasets that the object-centric model was
# trained on, they are simply skipped. These are for standard experiments - this field
# can be overridden to consider different variants, e.g. in the heldout experiments.
variant_types:
- original
- occlusion
- crop
- object_style
- object_color
- num_objects
- object_shape

output_features: all   # 'all' or list of strings

use_cache: true  # Cache representations in the first epoch when training downstream models

# The `overwrite` flag acts at the training set level. So we can run the downstream
# evaluation on new training sets, or using different downstream models or slot matching
# strategies and the results will be stored in new folders.
overwrite: false

debug: false

# Num workers for training only. Downstream training might be faster when this is 0.
num_workers: 0

hydra:
  run:
    dir: ./
  output_subdir: null

defaults:
  - override hydra/job_logging: disabled