config_pretext.yaml

batch_size: 64                  # training/evaluation batch size
epochs: 1                       # number of epochs to train. Ignored when running evaluation
validate_every_n_epochs: 1      # how often to run validation (epochs)
eval_every_n_epochs: 1          # how often to run evaluation (epochs)
fine_tune_from: None            # path to model to fine tune/evaluate
log_every_n_steps: 100          # how often to log to tensorboard (steps)
weight_decay: 10e-6             # training weight decay

model:                          # model parameters
  out_dim: 128                  # size of the projection output
  base_model: "scatsimclr30"    # base model to be used. Choices: "resnet18",  "resnet50", "scatsimclr8",  "scatsimclr12", "scatsimclr16", "scatsimclr30"
  J: 2                          # J - scale parameter of ScatNet. More: https://www.kymat.io/
  L: 16                         # L - rotation parameter of ScatNet. More: https://www.kymat.io/

dataset:                        # dataset parameters
  dataset: "stl10"              # dataset name: "stl10", "cifar20", "cifar10"
  input_shape: (96, 96, 3)      # input shape. For "stl10" - (96, 96, 3), for "cifar20", "cifar10" - (32, 32, 3)
  valid_size: 0.05              # percentage of valid data

loss:                          # loss parameters
  temperature: 0.5
  use_cosine_similarity: True

pretext:                       # pretext learning parameters
  jigsaw: False                # if True, jigsaw pretext task will be used. Only one of `jigsaw` or `rotation` can be True
  rotation: True               # if True, rotation pretext task will be used.
  num_jigsaw: 35               # number of jigsaw permutation. Ignored, when rotation. Note that each num_jigsaw requires file with permutations in data folder named permutations_{num_jigsaw}.npy
  lambda: 0.3                  # loss weight parameter for pretext task