Refactor eval.py (#127)

lerobot/configs/default.yaml

@@ -10,6 +10,8 @@ hydra:
     name: default
 
 device: cuda  # cpu
+# `seed` is used for training (eg: model initialization, dataset shuffling)
+# AND for the evaluation environments.
 seed: ???
 dataset_repo_id: lerobot/pusht
 
@@ -18,6 +20,8 @@ training:
   online_steps: ???
   online_steps_between_rollouts: ???
   online_sampling_ratio: 0.5
+  # `online_env_seed` is used for environments for online training data rollouts.
+  online_env_seed: ???
   eval_freq: ???
   save_freq: ???
   log_freq: 250
@@ -25,8 +29,10 @@ training:
 
 eval:
   n_episodes: 1
-  # TODO(alexander-soare): Right now this does not work. Reinstate this.
+  # `batch_size` specifies the number of environments to use in a gym.vector.VectorEnv.
   batch_size: 1
+  # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
+  use_async_envs: false
 
 wandb:
   enable: true

lerobot/configs/policy/act.yaml

@@ -28,7 +28,8 @@ training:
     action: "[i / ${fps} for i in range(${policy.chunk_size})]"
 
 eval:
-  n_episodes:: 50
+  n_episodes: 50
+  batch_size: 50
 
 # See `configuration_act.py` for more details.
 policy:

lerobot/configs/policy/diffusion.yaml

@@ -28,6 +28,7 @@ training:
 
 eval:
   n_episodes: 50
+  batch_size: 50
 
 override_dataset_stats:
   # TODO(rcadene, alexander-soare): should we remove image stats as well? do we use a pretrained vision model?

lerobot/configs/policy/tdmpc.yaml

@@ -8,6 +8,7 @@ training:
   eval_freq: 5000
   online_steps_between_rollouts: 1
   online_sampling_ratio: 0.5
+  online_env_seed: 10000
 
   batch_size: 256
   grad_clip_norm: 10.0