admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-05-22 20:19:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

nit

Browse Source
This commit is contained in:
Dana
2025-06-30 12:01:32 +02:00
parent c0146eed7f
commit 96550e4ad1
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
lerobot/common/policies/smolvla2/configuration_smolvla2.py
+4 -3
View File
@@ -74,11 +74,11 @@ class SmolVLAConfig(PreTrainedConfig):
# Finetuning settings # Finetuning settings
freeze_vision_encoder: bool = True freeze_vision_encoder: bool = True
train_expert_only: bool = True train_expert_only: bool = False
train_state_proj: bool = True train_state_proj: bool = True
# Training presets # 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_betas: tuple[float, float] = (0.9, 0.95)
optimizer_eps: float = 1e-8 optimizer_eps: float = 1e-8
optimizer_weight_decay: float = 1e-10 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_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 num_vlm_layers: int = 16
past_obs_keys: str = f"image"
add_local_special_image_tokens: bool = False add_local_special_image_tokens: bool = False
reverse_images_order: bool = False reverse_images_order: bool = False
@@ -115,7 +116,7 @@ class SmolVLAConfig(PreTrainedConfig):
causal_action_attention_mask: bool = False 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 = -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) 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 min_period: float = 4e-3 # sensitivity range for the timestep used in sine-cosine positional encoding