forked from tangger/lerobot
Add API Examples (#2289)
* (unscrewing things up) (#2288) * fix: expose a function explicitly building a frame for inference * fix: first make dataset frame, then make ready for inference * fix: reducing reliance on lerobot record for policy's ouptuts too * fix: encapsulating squeezing out + device handling from predict action * fix: remove duplicated call to build_inference_frame and add a function to only perform data type handling (whole conversion is: keys matching + data type conversion) * refactor(envs): add custom-observation-size (#2167) * fix: add MockMotorBus to MockRobot * rl: first drafts * add: all components of HIL SERL * fix: actor block works * fix: less friction, less friction * add: hil-serl complete example * fix: dataset names * fix: restructuring example folder * fix: act works but found bug in how ACT works * fix: same path for both pre and postprocessors * fix: paths * add: example usage for act * add: using ACT example * fix: training examples * fix: using examples * fix: camera index * fix: rename workflows into tutorial so that the path of the files is lerobot/examples/tutorial/... * fix: upload everything in one repo * fix: model name * fix: simplify model path * add: VLAs example --------- Signed-off-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com> * fix: minor fix using named attributes * fix: change model to act * fix: named attributes for inference frame building * fix: minor fixes to smolvla * fix: small changes to pi0 * remove: old file that should have never been committed (ups sorry sorry) --------- Signed-off-by: Francesco Capuano <74058581+fracapuano@users.noreply.github.com>
This commit is contained in:
Francesco Capuano
committed by
GitHub
GitHub
parent
be46bdea8f
commit
326aca0a48
99
examples/tutorial/diffusion/diffusion_training_example.py
Normal file
99
examples/tutorial/diffusion/diffusion_training_example.py
Normal file
@@ -0,0 +1,99 @@
|
||||
"""This script demonstrates how to train Diffusion Policy on a real-world dataset."""
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
import torch
|
||||
|
||||
from lerobot.configs.types import FeatureType
|
||||
from lerobot.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
|
||||
from lerobot.datasets.utils import dataset_to_policy_features
|
||||
from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
|
||||
from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
|
||||
from lerobot.policies.factory import make_pre_post_processors
|
||||
|
||||
|
||||
def make_delta_timestamps(delta_indices: list[int] | None, fps: int) -> list[float]:
|
||||
if delta_indices is None:
|
||||
return [0]
|
||||
|
||||
return [i / fps for i in delta_indices]
|
||||
|
||||
|
||||
output_directory = Path("outputs/robot_learning_tutorial/diffusion")
|
||||
output_directory.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
# Select your device
|
||||
device = torch.device("mps") # or "cuda" or "cpu"
|
||||
|
||||
dataset_id = "lerobot/svla_so101_pickplace"
|
||||
|
||||
# This specifies the inputs the model will be expecting and the outputs it will produce
|
||||
dataset_metadata = LeRobotDatasetMetadata(dataset_id)
|
||||
features = dataset_to_policy_features(dataset_metadata.features)
|
||||
|
||||
output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
|
||||
input_features = {key: ft for key, ft in features.items() if key not in output_features}
|
||||
|
||||
cfg = DiffusionConfig(input_features=input_features, output_features=output_features)
|
||||
policy = DiffusionPolicy(cfg)
|
||||
preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_metadata.stats)
|
||||
|
||||
policy.train()
|
||||
policy.to(device)
|
||||
|
||||
# To perform action chunking, ACT expects a given number of actions as targets
|
||||
delta_timestamps = {
|
||||
"observation.state": make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps),
|
||||
"action": make_delta_timestamps(cfg.action_delta_indices, dataset_metadata.fps),
|
||||
}
|
||||
|
||||
# add image features if they are present
|
||||
delta_timestamps |= {
|
||||
k: make_delta_timestamps(cfg.observation_delta_indices, dataset_metadata.fps) for k in cfg.image_features
|
||||
}
|
||||
|
||||
# Instantiate the dataset
|
||||
dataset = LeRobotDataset(dataset_id, delta_timestamps=delta_timestamps)
|
||||
|
||||
# Create the optimizer and dataloader for offline training
|
||||
optimizer = cfg.get_optimizer_preset().build(policy.parameters())
|
||||
batch_size = 32
|
||||
dataloader = torch.utils.data.DataLoader(
|
||||
dataset,
|
||||
batch_size=batch_size,
|
||||
shuffle=True,
|
||||
pin_memory=device.type != "cpu",
|
||||
drop_last=True,
|
||||
)
|
||||
|
||||
# Number of training steps and logging frequency
|
||||
training_steps = 1
|
||||
log_freq = 1
|
||||
|
||||
# Run training loop
|
||||
step = 0
|
||||
done = False
|
||||
while not done:
|
||||
for batch in dataloader:
|
||||
batch = preprocessor(batch)
|
||||
loss, _ = policy.forward(batch)
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
optimizer.zero_grad()
|
||||
|
||||
if step % log_freq == 0:
|
||||
print(f"step: {step} loss: {loss.item():.3f}")
|
||||
step += 1
|
||||
if step >= training_steps:
|
||||
done = True
|
||||
break
|
||||
|
||||
# Save the policy checkpoint, alongside the pre/post processors
|
||||
policy.save_pretrained(output_directory)
|
||||
preprocessor.save_pretrained(output_directory)
|
||||
postprocessor.save_pretrained(output_directory)
|
||||
|
||||
# Save all assets to the Hub
|
||||
policy.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
|
||||
preprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
|
||||
postprocessor.push_to_hub("fracapuano/robot_learning_tutorial_diffusion")
|
||||
60
examples/tutorial/diffusion/diffusion_using_example.py
Normal file
60
examples/tutorial/diffusion/diffusion_using_example.py
Normal file
@@ -0,0 +1,60 @@
|
||||
import torch
|
||||
|
||||
from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
|
||||
from lerobot.datasets.lerobot_dataset import LeRobotDatasetMetadata
|
||||
from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
|
||||
from lerobot.policies.factory import make_pre_post_processors
|
||||
from lerobot.policies.utils import build_inference_frame, make_robot_action
|
||||
from lerobot.robots.so100_follower.config_so100_follower import SO100FollowerConfig
|
||||
from lerobot.robots.so100_follower.so100_follower import SO100Follower
|
||||
|
||||
device = torch.device("mps") # or "cuda" or "cpu"
|
||||
model_id = "fracapuano/robot_learning_tutorial_diffusion"
|
||||
|
||||
model = DiffusionPolicy.from_pretrained(model_id)
|
||||
|
||||
dataset_id = "lerobot/svla_so101_pickplace"
|
||||
# This only downloads the metadata for the dataset, ~10s of MB even for large-scale datasets
|
||||
dataset_metadata = LeRobotDatasetMetadata(dataset_id)
|
||||
preprocess, postprocess = make_pre_post_processors(
|
||||
model.config, model_id, dataset_stats=dataset_metadata.stats
|
||||
)
|
||||
|
||||
MAX_EPISODES = 5
|
||||
MAX_STEPS_PER_EPISODE = 20
|
||||
|
||||
|
||||
# # find ports using lerobot-find-port
|
||||
follower_port = ... # something like "/dev/tty.usbmodem58760431631"
|
||||
|
||||
# # the robot ids are used the load the right calibration files
|
||||
follower_id = ... # something like "follower_so100"
|
||||
|
||||
# Robot and environment configuration
|
||||
# Camera keys must match the name and resolutions of the ones used for training!
|
||||
# You can check the camera keys expected by a model in the info.json card on the model card on the Hub
|
||||
camera_config = {
|
||||
"side": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
|
||||
"up": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
|
||||
}
|
||||
|
||||
robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_config)
|
||||
robot = SO100Follower(robot_cfg)
|
||||
robot.connect()
|
||||
|
||||
|
||||
for _ in range(MAX_EPISODES):
|
||||
for _ in range(MAX_STEPS_PER_EPISODE):
|
||||
obs = robot.get_observation()
|
||||
obs_frame = build_inference_frame(
|
||||
observation=obs, ds_features=dataset_metadata.features, device=device
|
||||
)
|
||||
|
||||
obs = preprocess(obs_frame)
|
||||
|
||||
action = model.select_action(obs)
|
||||
action = postprocess(action)
|
||||
action = make_robot_action(action, dataset_metadata.features)
|
||||
robot.send_action(action)
|
||||
|
||||
print("Episode finished! Starting new episode...")
|
||||
Reference in New Issue
Block a user