train_unet_ddim.yaml
======================

.. code-block:: yaml
  :linenos:

  # train ae or vae
  rand_seed: 2024
  model:
    # can be AE, VAE and DDPM
    name: DDIM
    num_steps: 1000
    sample_num_steps: 100
    beta_schedule: consine
    eps_model:
      # encoder config
      time_channel: 128
      encoder:
        name: UNet
        image_size: [28, 28]
        in_channel: 1
        patch_size: 1
        down_channels: [32, 64]
        middle_channels: [128, 128]
        building_block: res
        activation: silu
        num_blocks: 2
        normalization: group
        num_norm_groups: 16
        dropout: 0.1
      # condition_embedding:
      #   combine_condition: add
      #   merge_timestep_and_condition: true
      #   encoder:
      #     name: OneHot
      #     type: categories
      #     out_channel: 128
      #     num_classes: 10
    # classifier_free_guidance:
    #   condition_dropout: 0.2
    #   guidance_weight: 2.0
    ## 2D ddpm
    eps_loss:
      name: MSE
  loader:
    dataset: 
      name: MNIST
    data_path_list: 
      - /media/wlsdzyzl/DATA1/datasets/img/MNIST
    batch_size: 128
    num_workers: 8
    shuffle: true
    data_transforms:
      - name: Resize
        ## the value must be list
        size: [28, 28]
      - name: ToTensor
      - name: Normalize
        mean: [0.5]
        std: [0.5]
  check_point_dir: path/to/checkpoint/MNIST/DDPM
  sampler:
    num_sample_steps: 100
    rand_seed: 2024
    clipped: true
    clip_range: [-1.0, 1.0]
  ### parameter for optimizer
  optimizer:
    name: AdamW
    lr: 0.0003
    weight_decay: 0.00000001
  ### scheduler for learning rate
  lr_scheduler: 
    name: LinearLR
    start_factor: 1.0
    end_factor: 0.01
  max_epoch: 1000
  write_after_iters: 50
  save_after_epochs: 5