update: refactor configs and codes

scripts/configs/bt_awac/metaworld/state_dense.yaml

@@ -1,5 +1,5 @@
-algorithm: BTAWAC
-algorithm_kwargs:
+algorithm:
+  class: BTAWAC
   beta: 0.3333
   max_exp_clip: 100.0
   reward_steps: 50000
@@ -28,36 +28,31 @@ env_wrapper_kwargs:
 optim:
   default:
     class: Adam
-    kwargs:
-      lr: 0.0003
+    lr: 0.0003
 
 network:
   reward:
     class: EnsembleMLP
-    kwargs:
-      ensemble_size: 1
-      hidden_dims: [512, 512]
+    ensemble_size: 1
+    hidden_dims: [512, 512]
   actor:
     class: SquashedDeterministicActor
-    kwargs:
-      dropout: 0.25
-      hidden_dims: [512, 512]
+    dropout: 0.25
+    hidden_dims: [512, 512]
   critic:
     class: Critic
-    kwargs:
-      ensemble_size: 2
-      hidden_dims: [512, 512]
+    ensemble_size: 2
+    hidden_dims: [512, 512]
 
 dataset:
   class: MetaworldComparisonOfflineDataset
-  kwargs:
-    env: mw_drawer-open-v2
-    label_key: rl_sum
-    segment_length: null
-    batch_size: 96
-    capacity: 2500
-    mode: dense  # Choices: {sparse, dense}
-    discount: 0.99
+  env: mw_drawer-open-v2
+  label_key: rl_sum
+  segment_length: null
+  batch_size: 96
+  capacity: 2500
+  mode: dense  # Choices: {sparse, dense}
+  discount: 0.99
 dataloader:
   num_workers: 0  # use the main thread to sample data
   batch_size: null  # do not merge the data along batch axis
@@ -71,14 +66,12 @@ trainer:
 
 eval:
   function: eval_offline
-  kwargs:
-    num_ep: 10
-    deterministic: true
+  num_ep: 10
+  deterministic: true
 
 schedulers:
   actor:
     class: CosineAnnealingLR
-    kwargs:
-      T_max: 500000
+    T_max: 500000
 
 processor: null

scripts/configs/bt_awac/metaworld/state_sparse.yaml

@@ -1,5 +1,5 @@
-algorithm: BTAWAC
-algorithm_kwargs:
+algorithm:
+  class: BTAWAC
   beta: 0.3333
   max_exp_clip: 100.0
   reward_steps: 50000
@@ -28,36 +28,31 @@ env_wrapper_kwargs:
 optim:
   default:
     class: Adam
-    kwargs:
-      lr: 0.0003
+    lr: 0.0003
 
 network:
   reward:
     class: EnsembleMLP
-    kwargs:
-      ensemble_size: 1
-      hidden_dims: [512, 512]
+    ensemble_size: 1
+    hidden_dims: [512, 512]
   actor:
     class: SquashedDeterministicActor
-    kwargs:
-      dropout: 0.25
-      hidden_dims: [512, 512]
+    dropout: 0.25
+    hidden_dims: [512, 512]
   critic:
     class: Critic
-    kwargs:
-      ensemble_size: 2
-      hidden_dims: [512, 512]
+    ensemble_size: 2
+    hidden_dims: [512, 512]
 
 dataset:
   class: MetaworldComparisonOfflineDataset
-  kwargs:
-    env: mw_drawer-open-v2
-    label_key: rl_sum
-    segment_length: null
-    batch_size: 96
-    capacity: 10000
-    mode: sparse  # Choices: {sparse, dense}
-    discount: 0.99
+  env: mw_drawer-open-v2
+  label_key: rl_sum
+  segment_length: null
+  batch_size: 96
+  capacity: 10000
+  mode: sparse  # Choices: {sparse, dense}
+  discount: 0.99
 dataloader:
   num_workers: 0  # use the main thread to sample data
   batch_size: null  # do not merge the data along batch axis
@@ -71,14 +66,12 @@ trainer:
 
 eval:
   function: eval_offline
-  kwargs:
-    num_ep: 10
-    deterministic: true
+  num_ep: 10
+  deterministic: true
 
 schedulers:
   actor:
     class: CosineAnnealingLR
-    kwargs:
-      T_max: 500000
+    T_max: 500000
 
 processor: null

scripts/configs/bt_iql/gym/default.yaml

@@ -1,5 +1,5 @@
-algorithm: BTIQL
-algorithm_kwargs:
+algorithm:
+  class: BTIQL
   beta: 0.3333
   expectile: 0.7
   max_exp_clip: 100.0
@@ -24,46 +24,39 @@ env_wrapper_kwargs:
 optim:
   default:
     class: Adam
-    kwargs:
-      lr: 0.0003
+    lr: 0.0003
 
 network:
   reward:
     class: EnsembleMLP
-    kwargs:
-      ensemble_size: 1
-      hidden_dims: [256, 256]
+    ensemble_size: 1
+    hidden_dims: [256, 256]
   actor:
     class: SquashedGaussianActor
-    kwargs:
-      hidden_dims: [256, 256]
-      reparameterize: false
-      conditioned_logstd: false
-      logstd_min: -5
-      logstd_max: 2
+    hidden_dims: [256, 256]
+    reparameterize: false
+    conditioned_logstd: false
+    logstd_min: -5
+    logstd_max: 2
   critic:
     class: Critic
-    kwargs:
-      ensemble_size: 2
-      hidden_dims: [256, 256]
+    ensemble_size: 2
+    hidden_dims: [256, 256]
   value:
     class: Critic
-    kwargs:
-      ensemble_size: 1
-      hidden_dims: [256, 256]
+    ensemble_size: 1
+    hidden_dims: [256, 256]
 
 dataset:
   - class: IPLComparisonOfflineDataset
-    kwargs:
-      env: hopper-medium-replay-v2
-      batch_size: 8
-      segment_length: null
+    env: hopper-medium-replay-v2
+    batch_size: 8
+    segment_length: null
   - class: D4RLOfflineDataset
-    kwargs:
-      env: hopper-medium-replay-v2
-      batch_size: 256
-      mode: transition
-      reward_normalize: true
+    env: hopper-medium-replay-v2
+    batch_size: 256
+    mode: transition
+    reward_normalize: true
 
 dataloader:
   num_workers: 0  # use the main thread to sample data
@@ -78,14 +71,12 @@ trainer:
 
 eval:
   function: eval_offline
-  kwargs:
-    num_ep: 10
-    deterministic: true
+  num_ep: 10
+  deterministic: true
 
 schedulers:
   actor:
     class: CosineAnnealingLR
-    kwargs:
-      T_max: 1000000
+    T_max: 1000000
 
 processor: null

scripts/configs/bt_iql/metaworld/state_dense.yaml

@@ -1,5 +1,5 @@
-algorithm: BTIQL
-algorithm_kwargs:
+algorithm:
+  class: BTIQL
   beta: 0.3333
   expectile: 0.7
   max_exp_clip: 100.0
@@ -29,41 +29,35 @@ env_wrapper_kwargs:
 optim:
   default:
     class: Adam
-    kwargs:
-      lr: 0.0003
+    lr: 0.0003
 
 network:
   reward:
     class: EnsembleMLP
-    kwargs:
-      ensemble_size: 1
-      hidden_dims: [512, 512]
+    ensemble_size: 1
+    hidden_dims: [512, 512]
   actor:
     class: SquashedDeterministicActor
-    kwargs:
-      dropout: 0.25
-      hidden_dims: [512, 512]
+    dropout: 0.25
+    hidden_dims: [512, 512]
   critic:
     class: Critic
-    kwargs:
-      ensemble_size: 2
-      hidden_dims: [512, 512]
+    ensemble_size: 2
+    hidden_dims: [512, 512]
   value:
     class: Critic
-    kwargs:
-      ensemble_size: 1
-      hidden_dims: [512, 512]
+    ensemble_size: 1
+    hidden_dims: [512, 512]
 
 dataset:
   class: MetaworldComparisonOfflineDataset
-  kwargs:
-    env: mw_drawer-open-v2
-    label_key: rl_sum
-    segment_length: null
-    batch_size: 96
-    capacity: 2500
-    mode: dense  # Choices: {sparse, dense}
-    discount: 0.99
+  env: mw_drawer-open-v2
+  label_key: rl_sum
+  segment_length: null
+  batch_size: 96
+  capacity: 2500
+  mode: dense  # Choices: {sparse, dense}
+  discount: 0.99
 dataloader:
   num_workers: 0  # use the main thread to sample data
   batch_size: null  # do not merge the data along batch axis
@@ -77,14 +71,12 @@ trainer:
 
 eval:
   function: eval_offline
-  kwargs:
-    num_ep: 10
-    deterministic: true
+  num_ep: 10
+  deterministic: true
 
 schedulers:
   actor:
     class: CosineAnnealingLR
-    kwargs:
-      T_max: 500000
+    T_max: 500000
 
 processor: null