ijepa3/linear_eval_encoder_100/config.yaml

mode: evaluate

data:
  datasets_path: /home/lmk22/datasets
  train_dataset: imagenet
  batch_size: 1024 # For each device, so the effective batch size is batch_size * num_devices
  inference_batch_size: 1024 # For each device, so the effective batch size is inference_batch_size * num_devices
  crop_size: 224
  num_workers: 20 # For each device, so the effective number of workers is num_workers * num_devices
  prefetch_factor: 1
  pin_memory: false
  drop_last: true
  normalize:
    mean:
      - 0.485
      - 0.456
      - 0.406
    std:
      - 0.229
      - 0.224
      - 0.225
  horizontal_flip:
    use: true
    p: 0.5
  random_resized_crop:
    use: true
    p: 1.0
    scale:
    - 0.08
    - 1.0
    ratio:
    - 0.75
    - 1.3333333333333333
  central_crop_on_test:
    use: true
    initial_crop_size: 256
  separate_val_subset:
    use: false
    size: 0.1

meta:
  checkpoint: false
  mode: linear_eval
  bn_on_classifier: true
  pretrained_weights: null # Path to the pretrained weights file. Set to null if you don't want to load pretrained weights.
  save_every: 0 # Set to 0 if you don't want to save intermediate checkpoints during training
  framework: ijepa # Pass null if none of the frameworks was used (e.g., supervised_resnet50, random_resnet50, etc.)

optimization:
  ipe_scale: 1.0
  lr:
  - 0.1
  - 0.1
  - 0.0
  weight_decay:
  - 0
  - 0
  epochs: 50
  warmup_epochs: 0
  optimizer: sgd
  criterion: cross_entropy