Added caching function in the learner_server and modeling sac in order to limit the number of forward passes through the pretrained encoder when its frozen.

Added tensordict dependencies
Updated the version of torch and torchvision

Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>

lerobot/configs/env/maniskill_example.yaml

@@ -5,14 +5,14 @@ fps: 20
 env:
   name: maniskill/pushcube
   task:  PushCube-v1
-  image_size: 64
+  image_size: 128
   control_mode: pd_ee_delta_pose
   state_dim: 25
   action_dim: 7
   fps: ${fps}
   obs: rgb
   render_mode: rgb_array
-  render_size: 64
+  render_size: 128
   device: cuda
 
   reward_classifier:

lerobot/configs/policy/sac_maniskill.yaml

@@ -31,7 +31,7 @@ training:
   online_env_seed: 10000
   online_buffer_capacity: 1000000
   online_buffer_seed_size: 0
-  online_step_before_learning: 5000
+  online_step_before_learning: 500
   do_online_rollout_async: false
   policy_update_freq: 1
 
@@ -52,19 +52,16 @@ policy:
   n_action_steps: 1
 
   shared_encoder: true
-  vision_encoder_name: null
-  # vision_encoder_name: "helper2424/resnet10"
-  # freeze_vision_encoder: true
-  freeze_vision_encoder: false
+  vision_encoder_name: "helper2424/resnet10"
+  freeze_vision_encoder: true
   input_shapes:
     # # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
     observation.state: ["${env.state_dim}"]
-    observation.image: [3, 64, 64]
-    observation.image.2: [3, 64, 64]
+    observation.image: [3, 128, 128]
   output_shapes:
     action: [7]
   
-  camera_number: 2
+  camera_number: 1
 
   # Normalization / Unnormalization
   input_normalization_modes: null