model:
  target: models.Tiffusion.tiffusion.Tiffusion
  params:
    seq_length: 240
    feature_size: 3
    n_layer_enc: 6
    n_layer_dec: 4
    d_model: 128  # 4 X 16
    timesteps: 500 # diffusion timesteps
    sampling_timesteps: 500
    loss_type: 'l1'
    beta_schedule: 'cosine'
    n_heads: 8
    mlp_hidden_times: 4
    attn_pd: 0.0
    resid_pd: 0.0
    kernel_size: 1
    padding_size: 0
    control_signal: []
      # - classifier-based-sum-control
      # - classifier-free-sum-control
      # - range-wise-peak-control

solver:
  base_lr: 2.0e-5
  max_epochs: 2230 # 11150
  results_folder: ../../../data/ckpt_baseline
  gradient_accumulate_every: 2
  save_cycle: 223 # 1115  # max_epochs // 5
  ema:
    decay: 0.995
    update_interval: 10

  scheduler:
    target: engine.lr_sch.ReduceLROnPlateauWithWarmup
    params:
      factor: 0.65
      patience: 200
      min_lr: 1.0e-5
      threshold: 1.0e-1
      threshold_mode: rel
      warmup_lr: 8.0e-4
      warmup: 100
      verbose: False

dataloader:
  train_dataset:
    target: utils.data_utils.real_datasets.RevenueDataset
    params:
      name: revenue
      proportion: 0.8  # Set to rate < 1 if training conditional generation
      # data_root: ./Data/datasets/stock_data.csv
      data_root: ../../../data/daily.csv
      window: 240  # seq_length
      save2npy: True
      neg_one_to_one: True
      seed: 2024
      period: train

  test_dataset:
    target: utils.data_utils.real_datasets.RevenueDataset
    params:
      name: revenue
      proportion: 0.8  # rate
      data_root: ../../../data/daily.csv
      window: 240  # seq_length
      save2npy: True
      neg_one_to_one: True
      seed: 2024
      period: test
      style: separate
      # distribution: geometric
      distribution: uniform
      missing_ratio: 0.5

    coefficient: 1.0e-2
    step_size: 5.0e-2
    sampling_steps: 500

  batch_size: 64
  sample_size: 256
  shuffle: True