chore: remove unused code (#2062)

src/lerobot/policies/sac/configuration_sac.py

@@ -139,8 +139,6 @@ class SACConfig(PreTrainedConfig):
     # Training parameter
     # Number of steps for online training
     online_steps: int = 1000000
-    # Seed for the online environment
-    online_env_seed: int = 10000
     # Capacity of the online replay buffer
     online_buffer_capacity: int = 100000
     # Capacity of the offline replay buffer

src/lerobot/policies/sac/modeling_sac.py

@@ -1061,15 +1061,3 @@ class TanhMultivariateNormalDiag(TransformedDistribution):
             x = transform(x)
 
         return x
-
-
-def _convert_normalization_params_to_tensor(normalization_params: dict) -> dict:
-    converted_params = {}
-    for outer_key, inner_dict in normalization_params.items():
-        converted_params[outer_key] = {}
-        for key, value in inner_dict.items():
-            converted_params[outer_key][key] = torch.tensor(value)
-            if "image" in outer_key:
-                converted_params[outer_key][key] = converted_params[outer_key][key].view(3, 1, 1)
-
-    return converted_params

src/lerobot/policies/vqbet/configuration_vqbet.py

@@ -82,7 +82,6 @@ class VQBeTConfig(PreTrainedConfig):
         gpt_n_head: Number of headers of GPT
         gpt_hidden_dim: Size of hidden dimensions of GPT
         dropout: Dropout rate for GPT
-        mlp_hidden_dim: Size of hidden dimensions of offset header / bin prediction headers parts of VQ-BeT
         offset_loss_weight:  A constant that is multiplied to the offset loss
         primary_code_loss_weight: A constant that is multiplied to the primary code prediction loss
         secondary_code_loss_weight: A constant that is multiplied to the secondary code prediction loss
@@ -125,7 +124,6 @@ class VQBeTConfig(PreTrainedConfig):
     gpt_n_head: int = 8
     gpt_hidden_dim: int = 512
     dropout: float = 0.1
-    mlp_hidden_dim: int = 1024
     offset_loss_weight: float = 10000.0
     primary_code_loss_weight: float = 5.0
     secondary_code_loss_weight: float = 0.5

src/lerobot/policies/vqbet/vqbet_utils.py

@@ -231,16 +231,6 @@ class GPT(nn.Module):
             torch.nn.init.zeros_(module.bias)
             torch.nn.init.ones_(module.weight)
 
-    def crop_block_size(self, gpt_block_size):
-        # model surgery to decrease the block size if necessary
-        # e.g. we may load the GPT2 pretrained model checkpoint (block size 1024)
-        # but want to use a smaller block size for some smaller, simpler model
-        assert gpt_block_size <= self.config.gpt_block_size
-        self.config.gpt_block_size = gpt_block_size
-        self.transformer.wpe.weight = nn.Parameter(self.transformer.wpe.weight[:gpt_block_size])
-        for block in self.transformer.h:
-            block.attn.bias = block.attn.bias[:, :, :gpt_block_size, :gpt_block_size]
-
     def configure_parameters(self):
         """
         This long function is unfortunately doing something very simple and is being very defensive: