scripts/train_diffusion/cfg.yaml

dataset:
  dataset_name: "sevirlr"
  img_height: 128
  img_width: 128
  in_len: 7
  out_len: 6
  seq_len: 13
  plot_stride: 1
  interval_real_time: 10
  sample_mode: "sequent"
  stride: 6
  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: "2"
layout:
  in_len: 7
  out_len: 6
  in_step: &in_step 1
  out_step: &out_step 1
  in_out_diff: &in_out_diff 1
  img_height: 128
  img_width: 128
  data_channels: 1
  layout: "NTHWC"
optim:
  total_batch_size: 8
  micro_batch_size: 2
  seed: 0
  float32_matmul_precision: "high"
  method: "adamw"
  lr: 1.0e-3
  wd: 1.0e-5
  betas: [0.9, 0.999]
  gradient_clip_val: 1.0
  max_epochs: 2000
  loss_type: "l2"
  # 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/loss"
  # monitor: "valid_loss_epoch"
  early_stop: false
  early_stop_mode: "min"
  early_stop_patience: 100
  save_top_k: 3
logging:
  logging_prefix: "PreDiff"
  monitor_lr: true
  monitor_device: false
  track_grad_norm: -1
  use_wandb: false
  profiler: null
  save_npy: true
trainer:
  check_val_every_n_epoch: 50
  log_step_ratio: 0.001
  precision: 32
  find_unused_parameters: false
  num_sanity_val_steps: 2
eval:
  train_example_data_idx_list: [0, ]
  val_example_data_idx_list: [0, 16, 32, 48, 64, 72, 96, 108, 128]
  test_example_data_idx_list: [0, 16, 32, 48, 64, 72, 96, 108, 128]
  eval_example_only: true
  eval_aligned: true
  eval_unaligned: true
  num_samples_per_context: 1
  fs: 20
  label_offset: [-0.5, 0.5]
  label_avg_int: false
  fvd_features: 400
model:
  diffusion:
    data_shape: [6, 128, 128, 1]
    beta_schedule: "linear"
    use_ema: true
    log_every_t: 100
    clip_denoised: false
    linear_start: 1e-4
    linear_end: 2e-2
    cosine_s: 8e-3
    given_betas: null
    original_elbo_weight: 0.
    v_posterior: 0.
    l_simple_weight: 1.
    parameterization: "eps"
    learn_logvar: true
    logvar_init: 0.
    # latent diffusion
    latent_shape: [6, 16, 16, 64]
    cond_stage_model: "__is_first_stage__"
    num_timesteps_cond: null
    cond_stage_trainable: false
    cond_stage_forward: null
    scale_by_std: false
    scale_factor: 1.0
    latent_cond_shape: [7, 16, 16, 64]
  align:
    alignment_type: "avg_x"
    guide_scale: 50.0
    model_type: "cuboid"
    model_args:
      input_shape: [6, 16, 16, 64]
      out_channels: 1
      base_units: 128
      scale_alpha: 1.0
      depth: [1, 1]
      downsample: 2
      downsample_type: "patch_merge"
      block_attn_patterns: "axial"
      num_heads: 4
      attn_drop: 0.1
      proj_drop: 0.1
      ffn_drop: 0.1
      ffn_activation: "gelu"
      gated_ffn: false
      norm_layer: "layer_norm"
      use_inter_ffn: true
      hierarchical_pos_embed: false
      pos_embed_type: "t+h+w"
      padding_type: "zeros"
      checkpoint_level: 0
      use_relative_pos: true
      self_attn_use_final_proj: true
      # global vectors
      num_global_vectors: 0
      use_global_vector_ffn: true
      use_global_self_attn: false
      separate_global_qkv: false
      global_dim_ratio: 1
      # initialization
      attn_linear_init_mode: "0"
      ffn_linear_init_mode: "0"
      ffn2_linear_init_mode: "2"
      attn_proj_linear_init_mode: "2"
      conv_init_mode: "0"
      down_linear_init_mode: "0"
      global_proj_linear_init_mode: "2"
      norm_init_mode: "0"
      # timestep embedding for diffusion
      time_embed_channels_mult: 4
      time_embed_use_scale_shift_norm: false
      time_embed_dropout: 0.0
      # readout
      pool: "attention"
      readout_seq: true
      out_len: 6
    model_ckpt_path: "pretrained_sevirlr_alignment_avg_x_cuboid_v1.pt"
  latent_model:
    input_shape: [7, 16, 16, 64]
    target_shape: [6, 16, 16, 64]
    base_units: 256
    # block_units: null
    scale_alpha: 1.0
    num_heads: 4
    attn_drop: 0.1
    proj_drop: 0.1
    ffn_drop: 0.1
    # inter-attn downsample/upsample
    downsample: 2
    downsample_type: "patch_merge"
    upsample_type: "upsample"
    upsample_kernel_size: 3
    # cuboid attention
    depth: [4, 4]
    self_pattern: "axial"
    # global vectors
    num_global_vectors: 0
    use_dec_self_global: false
    dec_self_update_global: true
    use_dec_cross_global: false
    use_global_vector_ffn: false
    use_global_self_attn: true
    separate_global_qkv: true
    global_dim_ratio: 1
    # mise
    ffn_activation: "gelu"
    gated_ffn: false
    norm_layer: "layer_norm"
    padding_type: "zeros"
    pos_embed_type: "t+h+w"
    checkpoint_level: 0
    use_relative_pos: true
    self_attn_use_final_proj: true
    # initialization
    attn_linear_init_mode: "0"
    ffn_linear_init_mode: "0"
    ffn2_linear_init_mode: "2"
    attn_proj_linear_init_mode: "2"
    conv_init_mode: "0"
    down_up_linear_init_mode: "0"
    global_proj_linear_init_mode: "2"
    norm_init_mode: "0"
    # timestep embedding for diffusion
    time_embed_channels_mult: 4
    time_embed_use_scale_shift_norm: false
    time_embed_dropout: 0.0
    unet_res_connect: true
  vae:
    pretrained_ckpt_path: "pretrained_sevirlr_vae_8x8x64_v1_2.pt"
    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