J92 yaml

---

# continued training from 384x30-t60-4300 posted 2020-07-01

# T60 games used:
# 0-10k      2020-06-26 to 2020-07-07
# 10k-26k    2020-06-26 to 2020-07-12  
# 26k-40k    2020-07-07 to 2020-07-17
# 40k-55k    2020-07-12 to 2020-07-22
# 55k-70k    2020-07-17 to 2020-07-28  
# 70k-85k    2020-07022 to 2020-08-05
# 85k-100k next
# etc each round adding 2M + 0.2 games

##########################################################################
name: '384x30.J92.1'                     
gpu: 0                                 

dataset: 
  num_chunks_train:  4000000           
  num_chunks_test:    400000
  input_train: 'E:/T60_training_data/testJ92/train.1/*/'
  input_test:  'E:/T60_training_data/testJ92/test.1/*/'
  train_workers: 16
  test_workers: 8
  
training:
    total_steps: 85000                # terminate (total) steps, for batch 4096 <10k steps per million games

    batch_size: 4096                  
    num_batch_splits: 32             

    mask_legal_moves: true            
   
    renorm: true
    renorm_max_r: 4.0                 
    renorm_max_d: 5.0                 
    
    max_grad_norm: 3.5                # set higher than avg grad_norm but close
    
    swa: true
    swa_steps: 25
    swa_max_n: 10

    policy_loss_weight: 1.0           # weight of policy loss
    value_loss_weight:  0.7           # weight of value loss
    q_ratio: 0.0        
    moves_left_loss_weight: 1.0       # 
    
    test_steps: 1000                  # eval test set values after this many steps
    train_avg_report_steps: 50        # training reports average values after this many steps
    checkpoint_steps: 1000            # optional frequency for checkpointing before finish
    num_test_positions: 400000        # 

    shuffle_size: 500000              # size of the shuffle buffer
    
    lr_values: 
         - 0.002
         - 0.0006
         - 0.0006
    lr_boundaries:                     # list of boundaries in steps
         - 2000
         - 100000
        
    path: 'C:/networks/'               # network storage dir

model:                                 
  filters: 384
  residual_blocks: 30
  se_ratio: 12
  policy: 'convolution'                # should also be default
  value: 'wdl'                         # should also be default
  moves_left: 'v1'                     
  input_type: 'classic'
...```