nit

lerobot/common/policies/smolvla2/configuration_smolvla2.py

@@ -74,11 +74,11 @@ class SmolVLAConfig(PreTrainedConfig):
 
     # Finetuning settings
     freeze_vision_encoder: bool = True
-    train_expert_only: bool = True
+    train_expert_only: bool = False
     train_state_proj: bool = True
 
     # Training presets
-    optimizer_lr: float = 1e-4
+    optimizer_lr: float = 2.5e-5 #1e-4
     optimizer_betas: tuple[float, float] = (0.9, 0.95)
     optimizer_eps: float = 1e-8
     optimizer_weight_decay: float = 1e-10
@@ -105,6 +105,7 @@ class SmolVLAConfig(PreTrainedConfig):
 
     num_expert_layers: int = -1  # Less or equal to 0 is the default where the action expert has the same number of layers of VLM. Otherwise the expert have less layers.
     num_vlm_layers: int = 16  
+    past_obs_keys: str = f"image"
     add_local_special_image_tokens: bool = False
 
     reverse_images_order: bool = False
@@ -115,7 +116,7 @@ class SmolVLAConfig(PreTrainedConfig):
     causal_action_attention_mask: bool = False
 
     self_attn_every_n_layers: int = -1# Number of layers used in the VLM (first num_vlm_layers layers)
-    self_attn_every_n_layers: int = 2  # Interleave SA layers each self_attn_every_n_layers
+    #self_attn_every_n_layers: int = 2  # Interleave SA layers each self_attn_every_n_layers
     expert_width_multiplier: float = 0.75  # The action expert hidden size (wrt to the VLM)
 
     min_period: float = 4e-3  # sensitivity range for the timestep used in sine-cosine positional encoding