Spaces:
Sleeping
Sleeping
File size: 5,908 Bytes
63ed3a7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 |
import logging
import numpy as np
import torch
from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
from lerobot.common.utils.utils import init_logging
from .base_inference import BaseInferenceEngine
logger = logging.getLogger(__name__)
class DiffusionInferenceEngine(BaseInferenceEngine):
"""
Diffusion Policy inference engine.
Handles image preprocessing, joint normalization, and action prediction
for Diffusion Policy models using diffusion-based action generation.
"""
def __init__(
self,
policy_path: str,
camera_names: list[str],
use_custom_joint_names: bool = True,
device: str | None = None,
):
super().__init__(policy_path, camera_names, use_custom_joint_names, device)
# Diffusion-specific configuration
self.num_inference_steps = 10 # Number of diffusion steps
self.supports_language = (
False # Diffusion policies typically don't use language
)
async def load_policy(self):
"""Load the Diffusion policy from the specified path."""
logger.info(f"Loading Diffusion policy from: {self.policy_path}")
try:
# Initialize hydra config for LeRobot
init_logging()
# Load the Diffusion policy
self.policy = DiffusionPolicy.from_pretrained(self.policy_path)
self.policy.to(self.device)
self.policy.eval()
# Set up image transforms based on policy config
if hasattr(self.policy, "config"):
self._setup_image_transforms()
self.is_loaded = True
logger.info(f"✅ Diffusion policy loaded successfully on {self.device}")
except Exception as e:
logger.exception(f"Failed to load Diffusion policy from {self.policy_path}")
msg = f"Failed to load Diffusion policy: {e}"
raise RuntimeError(msg) from e
def _setup_image_transforms(self):
"""Set up image transforms based on the policy configuration."""
try:
# Get image size from policy config
config = self.policy.config
image_size = getattr(config, "image_size", 224)
# Create transforms for each camera
for camera_name in self.camera_names:
# Use policy-specific transforms if available
if hasattr(self.policy, "image_processor"):
self.image_transforms[camera_name] = self.policy.image_processor
else:
# Fall back to default transform
from torchvision import transforms
self.image_transforms[camera_name] = transforms.Compose([
transforms.Resize((image_size, image_size)),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
),
])
except Exception as e:
logger.warning(f"Could not set up image transforms: {e}. Using defaults.")
async def predict(
self, images: dict[str, np.ndarray], joint_positions: np.ndarray, **kwargs
) -> np.ndarray:
"""
Run Diffusion policy inference to predict actions.
Args:
images: Dictionary of {camera_name: rgb_image_array}
joint_positions: Current joint positions in LeRobot standard order
**kwargs: Additional arguments (unused for Diffusion)
Returns:
Array of predicted actions
"""
if not self.is_loaded:
msg = "Policy not loaded. Call load_policy() first."
raise RuntimeError(msg)
try:
# Preprocess inputs
processed_images = self.preprocess_images(images)
processed_joints = self.preprocess_joint_positions(joint_positions)
# Prepare batch inputs for Diffusion policy
batch = self._prepare_batch(processed_images, processed_joints)
# Run inference
with torch.no_grad():
action = self.policy.predict(batch)
# Convert to numpy
if isinstance(action, torch.Tensor):
action = action.cpu().numpy()
logger.debug(f"Diffusion policy predicted action shape: {action.shape}")
return action
except Exception as e:
logger.exception("Diffusion policy inference failed")
msg = f"Diffusion policy inference failed: {e}"
raise RuntimeError(msg) from e
def _prepare_batch(
self, images: dict[str, torch.Tensor], joints: torch.Tensor
) -> dict:
"""
Prepare batch inputs for Diffusion policy model.
Args:
images: Preprocessed images
joints: Preprocessed joint positions
Returns:
Batch dictionary for Diffusion policy model
"""
batch = {}
# Add images to batch
for camera_name, image_tensor in images.items():
# Add batch dimension if needed
if len(image_tensor.shape) == 3:
image_tensor = image_tensor.unsqueeze(0)
batch[f"observation.images.{camera_name}"] = image_tensor
# Add joint positions
if len(joints.shape) == 1:
joints = joints.unsqueeze(0)
batch["observation.state"] = joints
return batch
def get_model_info(self) -> dict:
"""Get Diffusion policy-specific model information."""
info = super().get_model_info()
info.update({
"policy_type": "diffusion",
"supports_language": self.supports_language,
"num_inference_steps": self.num_inference_steps,
})
return info
|