fix log_alpha in modeling_sac: change to nn.parameter

added pretrained vision model in policy

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

lerobot/common/policies/sac/modeling_sac.py

@@ -126,7 +126,7 @@ class SACPolicy(
         # TODO (azouitine): Handle the case where the temparameter is a fixed
         # TODO (michel-aractingi): Put the log_alpha in cuda by default because otherwise
         # it triggers "can't optimize a non-leaf Tensor"
-        self.log_alpha = torch.tensor([0.0], requires_grad=True, device=torch.device("mps"))
+        self.log_alpha = nn.Parameter(torch.tensor([0.0]))
         self.temperature = self.log_alpha.exp().item()
 
     def reset(self):
@@ -634,7 +634,7 @@ class PretrainedImageEncoder(nn.Module):
         """Set up CNN encoder"""
         from transformers import AutoModel
 
-        self.image_enc_layers = AutoModel.from_pretrained(config.vision_encoder_name)
+        self.image_enc_layers = AutoModel.from_pretrained(config.vision_encoder_name, trust_remote_code=True)
         # self.image_enc_layers.pooler = Identity()
 
         if hasattr(self.image_enc_layers.config, "hidden_sizes"):

lerobot/configs/env/so100_real.yaml

@@ -1,6 +1,6 @@
 # @package _global_
 
-fps: 30
+fps: 10
 
 env:
   name: real_world
@@ -26,6 +26,6 @@ env:
     joint_masking_action_space: [1, 1, 1, 1, 0, 0] # disable wrist and gripper
 
   reward_classifier:
-    pretrained_path: outputs/classifier/checkpoints/best/pretrained_model
+    pretrained_path:  outputs/classifier/13-02-random-sample-resnet10-frozen/checkpoints/best/pretrained_model
     config_path: lerobot/configs/policy/hilserl_classifier.yaml
     

lerobot/configs/policy/sac_real.yaml

@@ -8,7 +8,7 @@
 #   env.gym.obs_type=environment_state_agent_pos \
 
 seed: 1
-dataset_repo_id: null # aractingi/push_green_cube_hf_cropped_resized
+dataset_repo_id: aractingi/push_cube_square_light_offline_demo_cropped_resized
 
 training:
   # Offline training dataloader
@@ -52,7 +52,7 @@ policy:
   n_action_steps: 1
 
   shared_encoder: true
-  # vision_encoder_name: null
+  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?

lerobot/scripts/server/learner_server.py

@@ -411,7 +411,7 @@ def add_actor_information_and_train(
         next_observations = batch["next_state"]
         done = batch["done"]
 
-        assert_and_breakpoint(observations=observations, actions=actions, next_state=next_observations)
+        check_nan_in_transition(observations=observations, actions=actions, next_state=next_observations)
 
         with policy_lock:
             loss_critic = policy.compute_loss_critic(
@@ -533,7 +533,6 @@ def make_optimizers_and_scheduler(cfg, policy: nn.Module):
     optimizer_critic = torch.optim.Adam(
         params=policy.critic_ensemble.parameters(), lr=policy.config.critic_lr
     )
-    # We wrap policy log temperature in list because this is a torch tensor and not a nn.Module
     optimizer_temperature = torch.optim.Adam(params=[policy.log_alpha], lr=policy.config.critic_lr)
     lr_scheduler = None
     optimizers = {