%YAML 1.2
---
name: 'kb1-256x40'                      # ideally no spaces
gpu: 1                                  # gpu id to process on

dataset:
  num_chunks: 500000                    # newest nof chunks to parse
  train_ratio: 0.90                     # trainingset ratio
  input_train: '/home/joe/Downloads/train/'
  input_test: '/home/joe/Downloads/test/'
  #input: ''        # For a one-shot run with all data in one directory.

training:
    batch_size: 4096                    # batch_size
    num_batch_splits: 32                # training batch splits
    test_steps: 500                    # eval test set values after this many steps
    train_avg_report_steps: 50          # training reports its average values after this many steps.
    total_steps: 500                # terminate after these steps
    warmup_steps: 0                     # if global step is less than this, scale the current LR by
    checkpoint_steps: 500              # optional frequency for checkpointing before finish
    shuffle_size: 300000                # size of the shuffle buffer
    lr_cyclical: false
    lr_cyclical_steps: 10000
    lr_cyclical_max: .004
    lr_cyclical_min: .002
    lr_values:                          # list of learning rates
        - 0.002
    lr_boundaries:                      # list of boundaries
        - 1000000
    policy_loss_weight: 1.0             # weight of policy loss
    value_loss_weight: 1.0              # weight of value loss
    path: '/home/joe/Downloads/lczero-training/Networks/'    # network storage dir
    swa: False
    swa_max_n: 0
    renorm: True
    renorm_max_r: 1
    renorm_max_d: 0
    #renorm_momentum: .99
    q_ratio: 0
    max_grad_norm: 3
    mask_legal_moves: True

model:
  filters: 256
  residual_blocks: 40
  se_ratio: 8
  policy_channels: 128
  policy: convolution
  value: wdl
...