Tidy up yaml configs (#121)

lerobot/configs/policy/act.yaml

@@ -1,21 +1,34 @@
 # @package _global_
 
-offline_steps: 80000
-online_steps: 0
+seed: 1000
+dataset_repo_id: lerobot/aloha_sim_insertion_human
 
-eval_episodes: 1
-eval_freq: 10000
-save_freq: 100000
-log_freq: 250
+training:
+  offline_steps: 80000
+  online_steps: 0
+  eval_freq: 10000
+  save_freq: 100000
+  log_freq: 250
+  save_model: true
 
-n_obs_steps: 1
-# when temporal_agg=False, n_action_steps=horizon
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
 
-override_dataset_stats:
-  observation.images.top:
-    # stats from imagenet, since we use a pretrained vision model
-    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
-    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+  override_dataset_stats:
+    observation.images.top:
+      # stats from imagenet, since we use a pretrained vision model
+      mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+      std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes:: 50
 
 # See `configuration_act.py` for more details.
 policy:
@@ -24,7 +37,7 @@ policy:
   pretrained_model_path:
 
   # Input / output structure.
-  n_obs_steps: ${n_obs_steps}
+  n_obs_steps: 1
   chunk_size: 100 # chunk_size
   n_action_steps: 100
 
@@ -66,15 +79,3 @@ policy:
   # Training and loss computation.
   dropout: 0.1
   kl_weight: 10.0
-
-  # ---
-  # TODO(alexander-soare): Remove these from the policy config.
-  batch_size: 8
-  lr: 1e-5
-  lr_backbone: 1e-5
-  weight_decay: 1e-4
-  grad_clip_norm: 10
-  utd: 1
-
-  delta_timestamps:
-    action: "[i / ${fps} for i in range(${policy.chunk_size})]"

lerobot/configs/policy/diffusion.yaml

@@ -1,22 +1,33 @@
 # @package _global_
 
 seed: 100000
-horizon: 16
-n_obs_steps: 2
-n_action_steps: 8
-dataset_obs_steps: ${n_obs_steps}
-past_action_visible: False
-keypoint_visible_rate: 1.0
+dataset_repo_id: lerobot/pusht
 
-eval_episodes: 50
-eval_freq: 5000
-save_freq: 5000
-log_freq: 250
+training:
+  offline_steps: 200000
+  online_steps: 0
+  eval_freq: 5000
+  save_freq: 5000
+  log_freq: 250
+  save_model: true
 
-offline_steps: 200000
-online_steps: 0
+  batch_size: 64
+  grad_clip_norm: 10
+  lr: 1.0e-4
+  lr_scheduler: cosine
+  lr_warmup_steps: 500
+  adam_betas: [0.95, 0.999]
+  adam_eps: 1.0e-8
+  adam_weight_decay: 1.0e-6
+  online_steps_between_rollouts: 1
 
-offline_prioritized_sampler: true
+  delta_timestamps:
+    observation.image: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    observation.state: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    action: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1 - ${policy.n_obs_steps} + ${policy.horizon})]"
+
+eval:
+  n_episodes: 50
 
 override_dataset_stats:
   # TODO(rcadene, alexander-soare): should we remove image stats as well? do we use a pretrained vision model?
@@ -38,9 +49,9 @@ policy:
   pretrained_model_path:
 
   # Input / output structure.
-  n_obs_steps: ${n_obs_steps}
-  horizon: ${horizon}
-  n_action_steps: ${n_action_steps}
+  n_obs_steps: 2
+  horizon: 16
+  n_action_steps: 8
 
   input_shapes:
     # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
@@ -84,23 +95,9 @@ policy:
 
   # ---
   # TODO(alexander-soare): Remove these from the policy config.
-  batch_size: 64
-  grad_clip_norm: 10
-  lr: 1.0e-4
-  lr_scheduler: cosine
-  lr_warmup_steps: 500
-  adam_betas: [0.95, 0.999]
-  adam_eps: 1.0e-8
-  adam_weight_decay: 1.0e-6
-  utd: 1
   use_ema: true
   ema_update_after_step: 0
   ema_min_alpha: 0.0
   ema_max_alpha: 0.9999
   ema_inv_gamma: 1.0
   ema_power: 0.75
-
-  delta_timestamps:
-    observation.image: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1)]"
-    observation.state: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1)]"
-    action: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1 - ${n_obs_steps} + ${policy.horizon})]"

lerobot/configs/policy/tdmpc.yaml

@@ -54,7 +54,7 @@ policy:
   seed_steps: 0
   update_freq: 2
   tau: 0.01
-  utd: 1
+  online_steps_between_rollouts: 1
 
   # offline rl
   # dataset_dir: ???