Refactor policy config

lerobot/configs/default.yaml

@@ -1,10 +1,13 @@
 defaults:
   - _self_
   - env: simxarm
+  - policy: tdmpc
 
 hydra:
   run:
-    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}_${hydra.job.name}
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}_${env.name}_${policy.name}_${hydra.job.name}
+  job:
+    name: default
 
 seed: 1337
 device: cuda
@@ -20,82 +23,7 @@ fps: ???
 
 env: ???
 
-policy:
-  name: tdmpc
-
-  reward_scale: 1.0
-
-  # xarm_lift
-  train_steps: ${train_steps}
-  episode_length: ${env.episode_length}
-  discount: 0.9
-  modality: 'all'
-
-  # pixels
-  frame_stack: 1
-  num_channels: 32
-  img_size: ${env.image_size}
-  state_dim: ???
-  action_dim: ???
-
-  # planning
-  mpc: true
-  iterations: 6
-  num_samples: 512
-  num_elites: 50
-  mixture_coef: 0.1
-  min_std: 0.05
-  max_std: 2.0
-  temperature: 0.5
-  momentum: 0.1
-  uncertainty_cost: 1
-
-  # actor
-  log_std_min: -10
-  log_std_max: 2
-
-  # learning
-  batch_size: 256
-  max_buffer_size: 10000
-  horizon: 5
-  reward_coef: 0.5
-  value_coef: 0.1
-  consistency_coef: 20
-  rho: 0.5
-  kappa: 0.1
-  lr: 3e-4
-  std_schedule: ${policy.min_std}
-  horizon_schedule: ${policy.horizon}
-  per: true
-  per_alpha: 0.6
-  per_beta: 0.4
-  grad_clip_norm: 10
-  seed_steps: 0
-  update_freq: 2
-  tau: 0.01
-  utd: 1
-
-  # offline rl
-  # dataset_dir: ???
-  data_first_percent: 1.0
-  is_data_clip: true
-  data_clip_eps: 1e-5
-  expectile: 0.9
-  A_scaling: 3.0
-
-  # offline->online
-  offline_steps: 25000 # ${train_steps}/2
-  pretrained_model_path: ""
-  # pretrained_model_path: "/home/rcadene/code/fowm/logs/xarm_lift/all/default/2/models/offline.pt"
-  # pretrained_model_path: "/home/rcadene/code/fowm/logs/xarm_lift/all/default/2/models/final.pt"
-  balanced_sampling: true
-  demo_schedule: 0.5
-
-  # architecture
-  enc_dim: 256
-  num_q: 5
-  mlp_dim: 512
-  latent_dim: 50
+policy: ???
 
 wandb:
   enable: true

lerobot/configs/env/pusht.yaml

@@ -1,9 +1,5 @@
 # @package _global_
 
-hydra:
-  job:
-    name: pusht
-
 eval_episodes: 50
 eval_freq: 7500
 save_freq: 75000

lerobot/configs/env/simxarm.yaml

@@ -1,9 +1,5 @@
 # @package _global_
 
-hydra:
-  job:
-    name: simxarm
-
 eval_episodes: 20
 eval_freq: 1000
 save_freq: 10000

lerobot/configs/policy/tdmpc.yaml

@@ -0,0 +1,78 @@
+# @package _global_
+
+policy:
+  name: tdmpc
+
+  reward_scale: 1.0
+
+  # xarm_lift
+  train_steps: ${train_steps}
+  episode_length: ${env.episode_length}
+  discount: 0.9
+  modality: 'all'
+
+  # pixels
+  frame_stack: 1
+  num_channels: 32
+  img_size: ${env.image_size}
+  state_dim: ???
+  action_dim: ???
+
+  # planning
+  mpc: true
+  iterations: 6
+  num_samples: 512
+  num_elites: 50
+  mixture_coef: 0.1
+  min_std: 0.05
+  max_std: 2.0
+  temperature: 0.5
+  momentum: 0.1
+  uncertainty_cost: 1
+
+  # actor
+  log_std_min: -10
+  log_std_max: 2
+
+  # learning
+  batch_size: 256
+  max_buffer_size: 10000
+  horizon: 5
+  reward_coef: 0.5
+  value_coef: 0.1
+  consistency_coef: 20
+  rho: 0.5
+  kappa: 0.1
+  lr: 3e-4
+  std_schedule: ${policy.min_std}
+  horizon_schedule: ${policy.horizon}
+  per: true
+  per_alpha: 0.6
+  per_beta: 0.4
+  grad_clip_norm: 10
+  seed_steps: 0
+  update_freq: 2
+  tau: 0.01
+  utd: 1
+
+  # offline rl
+  # dataset_dir: ???
+  data_first_percent: 1.0
+  is_data_clip: true
+  data_clip_eps: 1e-5
+  expectile: 0.9
+  A_scaling: 3.0
+
+  # offline->online
+  offline_steps: 25000 # ${train_steps}/2
+  pretrained_model_path: ""
+  # pretrained_model_path: "/home/rcadene/code/fowm/logs/xarm_lift/all/default/2/models/offline.pt"
+  # pretrained_model_path: "/home/rcadene/code/fowm/logs/xarm_lift/all/default/2/models/final.pt"
+  balanced_sampling: true
+  demo_schedule: 0.5
+
+  # architecture
+  enc_dim: 256
+  num_q: 5
+  mlp_dim: 512
+  latent_dim: 50