deis_cifar.yml

uncond.py:
  cache_dir: #...
  checkpointing_steps: 2000
  clip_grad_norm: 0.0
  clip_sample: false
  dataset_config_name: null
  dataset_name: null
  diffusion_beta_linear_params:
  - 0.0001
  - 0.02
  diffusion_beta_schedule: linear
  diffusion_num_steps: 1000
  ema_decay: 0.9999
  eval_batch_size: 256
  exp_name: # ...
  infer: deissnode
  inference_num_steps: 6 # change
  learning_rate: 0.0001
  logger: tensorboard
  lr_scheduler: constant
  lr_warmup_steps: 0
  mixed_precision: fp16
  num_epochs: 2500
  num_samples: 50000
  output_dir: # ...
  pixel_resolution: 32
  prediction_type: epsilon
  reference_batch_path: # ...
  resume_from_checkpoint: null
  reverse_variance: recommended
  # sample_with: null
  sampler_order: 3
  saving_epochs: 50
  score_abs_Ls_path: ./cifar10_Ls_score_abs.pt
  clip_normalizer_steps: 5
  seed: 23
  timestep_spacing: quadratic
  train: ddpm
  train_batch_size: 256
  train_data_dir: # ...
  unet_json: ./arch/openai_guided_cifar_32.json
  use_ema: true