mnist_rot.yaml

dataset: 'mnist'
seed: 1337
exp_name: 'mnist_rotation'
variant_name: 'exp_1'
# When you want to have many variants
# of the same experiment

config: 'data/generator/config/mnist.json'
# config corresponding to the prior
logdir: 'logs/'
device: 'cuda'

# training
print_freq: 40
batch_size: 50
# need a high batch size for a good estimate of mmd
num_real_images: 200
max_epochs: 2000
epoch_length: 1000 #number of samples that constitute one epoch
train_reconstruction: true
freeze_encoder: true
reconstruction_epochs: 7
use_dist_loss: true
use_task_loss: true
moving_avg_alpha: 0.7 # moving_avg_alpha for baseline

# MMD
mmd_dims: [64, 192]
# sizes of layers of inception to use for MMD. Check 
# the inception file for possible values
mmd_resize_input: false

optim:
  lr: 0.001
  lr_decay: 200 # number of epochs to decay after
  lr_decay_gamma: 0.5 # gamma to decay
  weight_decay: 0.00001

weight:
  class: 0.1 # weight for class during reconstruction training
  dist_mmd: 100.0 # multiplier for mmd

task:
  val_root: 'data/datagen/mnist/target_rot/'
  # data corresponding to the target configuration
  # usually you would generate one small version
  # and one large version of the target
  # Use the small version while training and 
  # the large version to report final results
  # this is not included in this code for simplicity
  # but is easy to add by editing the test 
  # function in the task network to report two
  # accuracies, out of which one would be used to train
  device: 'cuda'
  print_freq: 100
  freeze_encoder: true
  batch_size: 8
  input_dim: [32, 32]
  epochs: 2
  dataset: 'mnist'