configurations/algorithm/df_base.yaml

defaults:
  - base_pytorch_algo

# dataset-dependent configurations
x_shape: ${dataset.observation_shape}
frame_stack: 1
frame_skip: 1
data_mean: ${dataset.data_mean}
data_std: ${dataset.data_std}
external_cond_dim: 0 #${dataset.action_dim}
context_frames: ${dataset.context_length}
# training hyperparameters
weight_decay: 1e-4
warmup_steps: 10000
optimizer_beta: [0.9, 0.999]
# diffusion-related
uncertainty_scale: 1
guidance_scale: 0.0
chunk_size: 1 # -1 for full trajectory diffusion, number to specify diffusion chunk size
scheduling_matrix: autoregressive
noise_level: random_all
causal: True

diffusion:
  # training
  objective: pred_x0
  beta_schedule: cosine
  schedule_fn_kwargs: {}
  clip_noise: 20.0
  use_snr: False
  use_cum_snr: False
  use_fused_snr: False
  snr_clip: 5.0
  cum_snr_decay: 0.98
  timesteps: 1000
  # sampling
  sampling_timesteps: 50 # fixme, numer of diffusion steps, should be increased
  ddim_sampling_eta: 1.0
  stabilization_level: 10
  # architecture
  architecture:
    network_size: 64