dataset:
  dataset_name: "sevirlr"
  img_height: 128
  img_width: 128
  in_len: 0
  out_len: 1
  seq_len: 1
  plot_stride: 1
  interval_real_time: 10
  sample_mode: "sequent"
  stride: 1
  layout: "NTHWC"
  start_date: null
  train_test_split_date: [2019, 6, 1]
  end_date: null
  val_ratio: 0.1
  metrics_mode: "0"
  metrics_list: ['csi', 'pod', 'sucr', 'bias']
  threshold_list: [16, 74, 133, 160, 181, 219]
  aug_mode: "1"
layout:
  layout: "NHWC"
optim:
  total_batch_size: 256
  micro_batch_size: 32
  float32_matmul_precision: "high"
  seed: 0
  method: "adam"
  lr: 5e-6
  betas: [0.5, 0.9]
  gradient_clip_val: 1.0
  max_epochs: 500
  # scheduler
  warmup_percentage: 0.1
  lr_scheduler_mode: "cosine"
  min_lr_ratio: 1.0e-3
  warmup_min_lr_ratio: 0.1
  # early stopping
  monitor: "val/total_loss"
  early_stop: true
  early_stop_mode: "min"
  early_stop_patience: 5
  save_top_k: 3
logging:
  logging_name: "vae_nhabe_training_downsample_200_b16"
  run_id: null
  logging_prefix: "NHABE_VAE_GAN_SEVIR-LR"
  monitor_lr: true
  monitor_device: false
  track_grad_norm: -1
  use_wandb: true
trainer:
  check_val_every_n_epoch: 5
  log_step_ratio: 0.001
  precision: 32
  find_unused_parameters: True
  num_sanity_val_steps: 2
eval:
  train_example_data_idx_list: [0, ]
  val_example_data_idx_list: [0, ]
  test_example_data_idx_list: [0, 16, 32, 48, 64, 72, 96, 108, 128]
  eval_example_only: false
  num_vis: 10
model:
  data_channels: 1
  down_block_types: ['DownEncoderBlock2D', 'DownEncoderBlock2D', 'DownEncoderBlock2D', 'DownEncoderBlock2D']
  in_channels: 1
  block_out_channels: [128, 256, 512, 512]  # downsample `len(block_out_channels) - 1` times
  act_fn: 'silu'
  latent_channels: 64
  up_block_types: ['UpDecoderBlock2D', 'UpDecoderBlock2D', 'UpDecoderBlock2D', 'UpDecoderBlock2D']
  norm_num_groups: 32
  layers_per_block: 2
  out_channels: 1
  loss:
    disc_start: 50001
    kl_weight: 1e-6
    disc_weight: 0.5
    perceptual_weight: 0.0  # SEVIR does not have RGB channels
    disc_in_channels: 1