admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-05-15 08:39:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

docs: update document in response to Simplify configs PR (#1596)

Browse Source 
* docs: update document input/output_shapes -> input/output_features

* fix inconsistent quote (suggested by copilot reviewer)

* docs: shapes => PolicyFeature

* docs: relfect normalization_mapping and remove outdated
This commit is contained in:
Hirokazu Ishida
2026-02-03 04:05:58 +09:00
committed by GitHub
parent b18cef2e26
commit 9c24a09665
5 changed files with 34 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/configs/policies.py
+6 -6
View File
@@ -45,12 +45,12 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC): # type: igno
Args:
n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
current step and additional steps going back).
input_shapes: A dictionary defining the shapes of the input data for the policy.
output_shapes: A dictionary defining the shapes of the output data for the policy.
input_normalization_modes: A dictionary with key representing the modality and the value specifies the
normalization mode to apply.
output_normalization_modes: Similar dictionary as `input_normalization_modes`, but to unnormalize to
the original scale.
input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
"""
n_obs_steps: int = 1
src/lerobot/policies/act/configuration_act.py
+7 -16
View File
@@ -28,7 +28,7 @@ class ACTConfig(PreTrainedConfig):
Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
The parameters you will most likely need to change are the ones which depend on the environment / sensors.
Those are: `input_shapes` and 'output_shapes`.
Those are: `input_features` and `output_features`.
Notes on the inputs and outputs:
- Either:
@@ -48,21 +48,12 @@ class ACTConfig(PreTrainedConfig):
This should be no greater than the chunk size. For example, if the chunk size size 100, you may
set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
environment, and throws the other 50 out.
input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
the input data name, and the value is a list indicating the dimensions of the corresponding data.
For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
include batch dimension or temporal dimension.
output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
the output data name, and the value is a list indicating the dimensions of the corresponding data.
For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
and the value specifies the normalization mode to apply. The two available modes are "mean_std"
which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
[-1, 1] range.
output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
original scale. Note that this is also used for normalizing the training targets.
input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
pretrained_backbone_weights: Pretrained weights from torchvision to initialize the backbone.
`None` means no pretrained weights.
src/lerobot/policies/diffusion/configuration_diffusion.py
+7 -16
View File
@@ -30,7 +30,7 @@ class DiffusionConfig(PreTrainedConfig):
Defaults are configured for training with PushT providing proprioceptive and single camera observations.
The parameters you will most likely need to change are the ones which depend on the environment / sensors.
Those are: `input_shapes` and `output_shapes`.
Those are: `input_features` and `output_features`.
Notes on the inputs and outputs:
- "observation.state" is required as an input key.
@@ -48,21 +48,12 @@ class DiffusionConfig(PreTrainedConfig):
horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
See `DiffusionPolicy.select_action` for more details.
input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
the input data name, and the value is a list indicating the dimensions of the corresponding data.
For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
include batch dimension or temporal dimension.
output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
the output data name, and the value is a list indicating the dimensions of the corresponding data.
For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
and the value specifies the normalization mode to apply. The two available modes are "mean_std"
which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
[-1, 1] range.
output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
original scale. Note that this is also used for normalizing the training targets.
input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
within the image size. If None, no cropping is done.
src/lerobot/policies/tdmpc/configuration_tdmpc.py
+7 -19
View File
@@ -30,7 +30,7 @@ class TDMPCConfig(PreTrainedConfig):
camera observations.
The parameters you will most likely need to change are the ones which depend on the environment / sensors.
Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
Those are: `input_features`, `output_features`, and perhaps `max_random_shift_ratio`.
Args:
n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
@@ -40,24 +40,12 @@ class TDMPCConfig(PreTrainedConfig):
is an alternative to using action repeats. If this is set to more than 1, then we require
`n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
approach of using multiple steps from the plan is not in the original implementation.
input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
the input data name, and the value is a list indicating the dimensions of the corresponding data.
For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
indicating it has three color channels and 96x96 resolution. Importantly, `input_shapes` doesn't
include batch dimension or temporal dimension.
output_shapes: A dictionary defining the shapes of the output data for the policy. The key represents
the output data name, and the value is a list indicating the dimensions of the corresponding data.
For example, "action" refers to an output shape of [14], indicating 14-dimensional actions.
Importantly, `output_shapes` doesn't include batch dimension or temporal dimension.
input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
and the value specifies the normalization mode to apply. The two available modes are "mean_std"
which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
[-1, 1] range. Note that here this defaults to None meaning inputs are not normalized. This is to
match the original implementation.
output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
original scale. Note that this is also used for normalizing the training targets. NOTE: Clipping
to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
normalization mode here.
input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
latent_dim: Observation's latent embedding dimension.
src/lerobot/policies/vqbet/configuration_vqbet.py
+7 -16
View File
@@ -32,7 +32,7 @@ class VQBeTConfig(PreTrainedConfig):
Defaults are configured for training with PushT providing proprioceptive and single camera observations.
The parameters you will most likely need to change are the ones which depend on the environment / sensors.
Those are: `input_shapes` and `output_shapes`.
Those are: `input_features` and `output_features`.
Notes on the inputs and outputs:
- "observation.state" is required as an input key.
@@ -46,21 +46,12 @@ class VQBeTConfig(PreTrainedConfig):
current step and additional steps going back).
n_action_pred_token: Total number of current token and future tokens that VQ-BeT predicts.
action_chunk_size: Action chunk size of each action prediction token.
input_shapes: A dictionary defining the shapes of the input data for the policy.
The key represents the input data name, and the value is a list indicating the dimensions
of the corresponding data. For example, "observation.image" refers to an input from
a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
Importantly, shapes doesnt include batch dimension or temporal dimension.
output_shapes: A dictionary defining the shapes of the output data for the policy.
The key represents the output data name, and the value is a list indicating the dimensions
of the corresponding data. For example, "action" refers to an output shape of [14], indicating
14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
and the value specifies the normalization mode to apply. The two available modes are "mean_std"
which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
[-1, 1] range.
output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
original scale. Note that this is also used for normalizing the training targets.
input_features: A dictionary defining the PolicyFeature of the input data for the policy. The key represents
the input data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
output_features: A dictionary defining the PolicyFeature of the output data for the policy. The key represents
the output data name, and the value is PolicyFeature, which consists of FeatureType and shape attributes.
normalization_mapping: A dictionary that maps from a str value of FeatureType (e.g., "STATE", "VISUAL") to
a corresponding NormalizationMode (e.g., NormalizationMode.MIN_MAX)
vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
within the image size. If None, no cropping is done.