tangger/lerobot
1
0
Fork 1
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Simplify configs (#550)

Browse Source 
Co-authored-by: Remi <remi.cadene@huggingface.co>
Co-authored-by: HUANG TZU-CHUN <137322177+tc-huang@users.noreply.github.com>
This commit is contained in:
Simon Alibert
2025-01-31 13:57:37 +01:00
committed by GitHub
parent 1ee1acf8ad
commit 3c0a209f9f
119 changed files with 5761 additions and 5466 deletions
Download Patch File Download Diff File Expand all files Collapse all files

114
lerobot/common/policies/diffusion/configuration_diffusion.py
View File

@@ -16,9 +16,15 @@
# limitations under the License.
from dataclasses import dataclass, field
from lerobot.common.optim.optimizers import AdamConfig
from lerobot.common.optim.schedulers import DiffuserSchedulerConfig
from lerobot.configs.policies import PreTrainedConfig
from lerobot.configs.types import NormalizationMode
@PreTrainedConfig.register_subclass("diffusion")
@dataclass
class DiffusionConfig:
class DiffusionConfig(PreTrainedConfig):
"""Configuration class for DiffusionPolicy.
Defaults are configured for training with PushT providing proprioceptive and single camera observations.
@@ -102,26 +108,17 @@ class DiffusionConfig:
horizon: int = 16
n_action_steps: int = 8
input_shapes: dict[str, list[int]] = field(
normalization_mapping: dict[str, NormalizationMode] = field(
default_factory=lambda: {
"observation.image": [3, 96, 96],
"observation.state": [2],
}
)
output_shapes: dict[str, list[int]] = field(
default_factory=lambda: {
"action": [2],
"VISUAL": NormalizationMode.MEAN_STD,
"STATE": NormalizationMode.MIN_MAX,
"ACTION": NormalizationMode.MIN_MAX,
}
)
# Normalization / Unnormalization
input_normalization_modes: dict[str, str] = field(
default_factory=lambda: {
"observation.image": "mean_std",
"observation.state": "min_max",
}
)
output_normalization_modes: dict[str, str] = field(default_factory=lambda: {"action": "min_max"})
# The original implementation doesn't sample frames for the last 7 steps,
# which avoids excessive padding and leads to improved training results.
drop_n_last_frames: int = 7 # horizon - n_action_steps - n_obs_steps + 1
# Architecture / modeling.
# Vision backbone.
@@ -154,39 +151,23 @@ class DiffusionConfig:
# Loss computation
do_mask_loss_for_padding: bool = False
# Training presets
optimizer_lr: float = 1e-4
optimizer_betas: tuple = (0.95, 0.999)
optimizer_eps: float = 1e-8
optimizer_weight_decay: float = 1e-6
scheduler_name: str = "cosine"
scheduler_warmup_steps: int = 500
def __post_init__(self):
super().__post_init__()
"""Input validation (not exhaustive)."""
if not self.vision_backbone.startswith("resnet"):
raise ValueError(
f"`vision_backbone` must be one of the ResNet variants. Got {self.vision_backbone}."
)
image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
if len(image_keys) == 0 and "observation.environment_state" not in self.input_shapes:
raise ValueError("You must provide at least one image or the environment state among the inputs.")
if len(image_keys) > 0:
if self.crop_shape is not None:
for image_key in image_keys:
if (
self.crop_shape[0] > self.input_shapes[image_key][1]
or self.crop_shape[1] > self.input_shapes[image_key][2]
):
raise ValueError(
f"`crop_shape` should fit within `input_shapes[{image_key}]`. Got {self.crop_shape} "
f"for `crop_shape` and {self.input_shapes[image_key]} for "
"`input_shapes[{image_key}]`."
)
# Check that all input images have the same shape.
first_image_key = next(iter(image_keys))
for image_key in image_keys:
if self.input_shapes[image_key] != self.input_shapes[first_image_key]:
raise ValueError(
f"`input_shapes[{image_key}]` does not match `input_shapes[{first_image_key}]`, but we "
"expect all image shapes to match."
)
supported_prediction_types = ["epsilon", "sample"]
if self.prediction_type not in supported_prediction_types:
raise ValueError(
@@ -207,3 +188,50 @@ class DiffusionConfig:
"The horizon should be an integer multiple of the downsampling factor (which is determined "
f"by `len(down_dims)`). Got {self.horizon=} and {self.down_dims=}"
)
def get_optimizer_preset(self) -> AdamConfig:
return AdamConfig(
lr=self.optimizer_lr,
betas=self.optimizer_betas,
eps=self.optimizer_eps,
weight_decay=self.optimizer_weight_decay,
)
def get_scheduler_preset(self) -> DiffuserSchedulerConfig:
return DiffuserSchedulerConfig(
name=self.scheduler_name,
num_warmup_steps=self.scheduler_warmup_steps,
)
def validate_features(self) -> None:
if len(self.image_features) == 0 and self.env_state_feature is None:
raise ValueError("You must provide at least one image or the environment state among the inputs.")
if self.crop_shape is not None:
for key, image_ft in self.image_features.items():
if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
raise ValueError(
f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
f"for `crop_shape` and {image_ft.shape} for "
f"`{key}`."
)
# Check that all input images have the same shape.
first_image_key, first_image_ft = next(iter(self.image_features.items()))
for key, image_ft in self.image_features.items():
if image_ft.shape != first_image_ft.shape:
raise ValueError(
f"`{key}` does not match `{first_image_key}`, but we " "expect all image shapes to match."
)
@property
def observation_delta_indices(self) -> list:
return list(range(1 - self.n_obs_steps, 1))
@property
def action_delta_indices(self) -> list:
return list(range(1 - self.n_obs_steps, 1 - self.n_obs_steps + self.horizon))
@property
def reward_delta_indices(self) -> None:
return None

102
lerobot/common/policies/diffusion/modeling_diffusion.py
View File

@@ -31,35 +31,32 @@ import torch.nn.functional as F # noqa: N812
import torchvision
from diffusers.schedulers.scheduling_ddim import DDIMScheduler
from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
from huggingface_hub import PyTorchModelHubMixin
from torch import Tensor, nn
from lerobot.common.constants import OBS_ENV, OBS_ROBOT
from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
from lerobot.common.policies.normalize import Normalize, Unnormalize
from lerobot.common.policies.pretrained import PreTrainedPolicy
from lerobot.common.policies.utils import (
get_device_from_parameters,
get_dtype_from_parameters,
get_output_shape,
populate_queues,
)
class DiffusionPolicy(
nn.Module,
PyTorchModelHubMixin,
library_name="lerobot",
repo_url="https://github.com/huggingface/lerobot",
tags=["robotics", "diffusion-policy"],
):
class DiffusionPolicy(PreTrainedPolicy):
"""
Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
(paper: https://arxiv.org/abs/2303.04137, code: https://github.com/real-stanford/diffusion_policy).
"""
config_class = DiffusionConfig
name = "diffusion"
def __init__(
self,
config: DiffusionConfig | None = None,
config: DiffusionConfig,
dataset_stats: dict[str, dict[str, Tensor]] | None = None,
):
"""
@@ -69,18 +66,16 @@ class DiffusionPolicy(
dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
that they will be passed with a call to `load_state_dict` before the policy is used.
"""
super().__init__()
if config is None:
config = DiffusionConfig()
super().__init__(config)
config.validate_features()
self.config = config
self.normalize_inputs = Normalize(
config.input_shapes, config.input_normalization_modes, dataset_stats
)
self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
self.normalize_targets = Normalize(
config.output_shapes, config.output_normalization_modes, dataset_stats
config.output_features, config.normalization_mapping, dataset_stats
)
self.unnormalize_outputs = Unnormalize(
config.output_shapes, config.output_normalization_modes, dataset_stats
config.output_features, config.normalization_mapping, dataset_stats
)
# queues are populated during rollout of the policy, they contain the n latest observations and actions
@@ -88,20 +83,20 @@ class DiffusionPolicy(
self.diffusion = DiffusionModel(config)
self.expected_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
self.use_env_state = "observation.environment_state" in config.input_shapes
self.reset()
def get_optim_params(self) -> dict:
return self.diffusion.parameters()
def reset(self):
"""Clear observation and action queues. Should be called on `env.reset()`"""
self._queues = {
"observation.state": deque(maxlen=self.config.n_obs_steps),
"action": deque(maxlen=self.config.n_action_steps),
}
if len(self.expected_image_keys) > 0:
if self.config.image_features:
self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
if self.use_env_state:
if self.config.env_state_feature:
self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
@torch.no_grad
@@ -127,9 +122,11 @@ class DiffusionPolicy(
actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
"""
batch = self.normalize_inputs(batch)
if len(self.expected_image_keys) > 0:
if self.config.image_features:
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
batch["observation.images"] = torch.stack(
[batch[key] for key in self.config.image_features], dim=-4
)
# Note: It's important that this happens after stacking the images into a single key.
self._queues = populate_queues(self._queues, batch)
@@ -149,9 +146,11 @@ class DiffusionPolicy(
def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
"""Run the batch through the model and compute the loss for training or validation."""
batch = self.normalize_inputs(batch)
if len(self.expected_image_keys) > 0:
if self.config.image_features:
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack([batch[k] for k in self.expected_image_keys], dim=-4)
batch["observation.images"] = torch.stack(
[batch[key] for key in self.config.image_features], dim=-4
)
batch = self.normalize_targets(batch)
loss = self.diffusion.compute_loss(batch)
return {"loss": loss}
@@ -176,12 +175,9 @@ class DiffusionModel(nn.Module):
self.config = config
# Build observation encoders (depending on which observations are provided).
global_cond_dim = config.input_shapes["observation.state"][0]
num_images = len([k for k in config.input_shapes if k.startswith("observation.image")])
self._use_images = False
self._use_env_state = False
if num_images > 0:
self._use_images = True
global_cond_dim = self.config.robot_state_feature.shape[0]
if self.config.image_features:
num_images = len(self.config.image_features)
if self.config.use_separate_rgb_encoder_per_camera:
encoders = [DiffusionRgbEncoder(config) for _ in range(num_images)]
self.rgb_encoder = nn.ModuleList(encoders)
@@ -189,9 +185,8 @@ class DiffusionModel(nn.Module):
else:
self.rgb_encoder = DiffusionRgbEncoder(config)
global_cond_dim += self.rgb_encoder.feature_dim * num_images
if "observation.environment_state" in config.input_shapes:
self._use_env_state = True
global_cond_dim += config.input_shapes["observation.environment_state"][0]
if self.config.env_state_feature:
global_cond_dim += self.config.env_state_feature.shape[0]
self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
@@ -220,7 +215,7 @@ class DiffusionModel(nn.Module):
# Sample prior.
sample = torch.randn(
size=(batch_size, self.config.horizon, self.config.output_shapes["action"][0]),
size=(batch_size, self.config.horizon, self.config.action_feature.shape[0]),
dtype=dtype,
device=device,
generator=generator,
@@ -242,10 +237,10 @@ class DiffusionModel(nn.Module):
def _prepare_global_conditioning(self, batch: dict[str, Tensor]) -> Tensor:
"""Encode image features and concatenate them all together along with the state vector."""
batch_size, n_obs_steps = batch["observation.state"].shape[:2]
global_cond_feats = [batch["observation.state"]]
batch_size, n_obs_steps = batch[OBS_ROBOT].shape[:2]
global_cond_feats = [batch[OBS_ROBOT]]
# Extract image features.
if self._use_images:
if self.config.image_features:
if self.config.use_separate_rgb_encoder_per_camera:
# Combine batch and sequence dims while rearranging to make the camera index dimension first.
images_per_camera = einops.rearrange(batch["observation.images"], "b s n ... -> n (b s) ...")
@@ -272,8 +267,8 @@ class DiffusionModel(nn.Module):
)
global_cond_feats.append(img_features)
if self._use_env_state:
global_cond_feats.append(batch["observation.environment_state"])
if self.config.env_state_feature:
global_cond_feats.append(batch[OBS_ENV])
# Concatenate features then flatten to (B, global_cond_dim).
return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)
@@ -443,7 +438,7 @@ class SpatialSoftmax(nn.Module):
class DiffusionRgbEncoder(nn.Module):
"""Encoder an RGB image into a 1D feature vector.
"""Encodes an RGB image into a 1D feature vector.
Includes the ability to normalize and crop the image first.
"""
@@ -482,19 +477,16 @@ class DiffusionRgbEncoder(nn.Module):
# Set up pooling and final layers.
# Use a dry run to get the feature map shape.
# The dummy input should take the number of image channels from `config.input_shapes` and it should
# The dummy input should take the number of image channels from `config.image_features` and it should
# use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
# height and width from `config.input_shapes`.
image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
# height and width from `config.image_features`.
# Note: we have a check in the config class to make sure all images have the same shape.
image_key = image_keys[0]
dummy_input_h_w = (
config.crop_shape if config.crop_shape is not None else config.input_shapes[image_key][1:]
)
dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *dummy_input_h_w))
with torch.inference_mode():
dummy_feature_map = self.backbone(dummy_input)
feature_map_shape = tuple(dummy_feature_map.shape[1:])
images_shape = next(iter(config.image_features.values())).shape
dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
self.feature_dim = config.spatial_softmax_num_keypoints * 2
self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
@@ -611,7 +603,7 @@ class DiffusionConditionalUnet1d(nn.Module):
# In channels / out channels for each downsampling block in the Unet's encoder. For the decoder, we
# just reverse these.
in_out = [(config.output_shapes["action"][0], config.down_dims[0])] + list(
in_out = [(config.action_feature.shape[0], config.down_dims[0])] + list(
zip(config.down_dims[:-1], config.down_dims[1:], strict=True)
)
@@ -666,7 +658,7 @@ class DiffusionConditionalUnet1d(nn.Module):
self.final_conv = nn.Sequential(
DiffusionConv1dBlock(config.down_dims[0], config.down_dims[0], kernel_size=config.kernel_size),
nn.Conv1d(config.down_dims[0], config.output_shapes["action"][0], 1),
nn.Conv1d(config.down_dims[0], config.action_feature.shape[0], 1),
)
def forward(self, x: Tensor, timestep: Tensor | int, global_cond=None) -> Tensor: