hangX/lerobot
1
0
Fork 0
forked from tangger/lerobot
Code Pull Requests Activity

refactor(config): Move device & amp args to PreTrainedConfig (#812)

Browse Source 
Co-authored-by: Simon Alibert <75076266+aliberts@users.noreply.github.com>
This commit is contained in:
Steven Palma
2025-03-06 17:59:28 +01:00
committed by GitHub
parent 10706ed753
commit 5e9473806c
19 changed files with 62 additions and 136 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
lerobot/configs/eval.py
View File

@@ -18,11 +18,9 @@ from dataclasses import dataclass, field
from pathlib import Path
from lerobot.common import envs, policies # noqa: F401
from lerobot.common.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
from lerobot.configs import parser
from lerobot.configs.default import EvalConfig
from lerobot.configs.policies import PreTrainedConfig
from lerobot.configs.train import TrainPipelineConfig
@dataclass
@@ -35,11 +33,6 @@ class EvalPipelineConfig:
policy: PreTrainedConfig | None = None
output_dir: Path | None = None
job_name: str | None = None
# TODO(rcadene, aliberts): By default, use device and use_amp values from policy checkpoint.
device: str | None = None # cuda | cpu | mps
# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
# automatic gradient scaling is used.
use_amp: bool = False
seed: int | None = 1000
def __post_init__(self):
@@ -50,27 +43,6 @@ class EvalPipelineConfig:
self.policy = PreTrainedConfig.from_pretrained(policy_path, cli_overrides=cli_overrides)
self.policy.pretrained_path = policy_path
# When no device or use_amp are given, use the one from training config.
if self.device is None or self.use_amp is None:
train_cfg = TrainPipelineConfig.from_pretrained(policy_path)
if self.device is None:
self.device = train_cfg.device
if self.use_amp is None:
self.use_amp = train_cfg.use_amp
# Automatically switch to available device if necessary
if not is_torch_device_available(self.device):
auto_device = auto_select_torch_device()
logging.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
self.device = auto_device
# Automatically deactivate AMP if necessary
if self.use_amp and not is_amp_available(self.device):
logging.warning(
f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
)
self.use_amp = False
else:
logging.warning(
"No pretrained path was provided, evaluated policy will be built from scratch (random weights)."
@@ -87,11 +59,6 @@ class EvalPipelineConfig:
eval_dir = f"{now:%Y-%m-%d}/{now:%H-%M-%S}_{self.job_name}"
self.output_dir = Path("outputs/eval") / eval_dir
if self.device is None:
raise ValueError("Set one of the following device: cuda, cpu or mps")
elif self.device == "cuda" and self.use_amp is None:
raise ValueError("Set 'use_amp' to True or False.")
@classmethod
def __get_path_fields__(cls) -> list[str]:
"""This enables the parser to load config from the policy using `--policy.path=local/dir`"""

18
lerobot/configs/policies.py
View File

@@ -12,6 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import abc
import logging
import os
from dataclasses import dataclass, field
from pathlib import Path
@@ -25,6 +26,7 @@ from huggingface_hub.errors import HfHubHTTPError
from lerobot.common.optim.optimizers import OptimizerConfig
from lerobot.common.optim.schedulers import LRSchedulerConfig
from lerobot.common.utils.hub import HubMixin
from lerobot.common.utils.utils import auto_select_torch_device, is_amp_available, is_torch_device_available
from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
# Generic variable that is either PreTrainedConfig or a subclass thereof
@@ -53,8 +55,24 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):
input_features: dict[str, PolicyFeature] = field(default_factory=dict)
output_features: dict[str, PolicyFeature] = field(default_factory=dict)
device: str | None = None # cuda | cpu | mp
# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
# automatic gradient scaling is used.
use_amp: bool = False
def __post_init__(self):
self.pretrained_path = None
if not self.device or not is_torch_device_available(self.device):
auto_device = auto_select_torch_device()
logging.warning(f"Device '{self.device}' is not available. Switching to '{auto_device}'.")
self.device = auto_device.type
# Automatically deactivate AMP if necessary
if self.use_amp and not is_amp_available(self.device):
logging.warning(
f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
)
self.use_amp = False
@property
def type(self) -> str:

18
lerobot/configs/train.py
View File

@@ -12,7 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import datetime as dt
import logging
import os
from dataclasses import dataclass, field
from pathlib import Path
@@ -26,7 +25,6 @@ from lerobot.common import envs
from lerobot.common.optim import OptimizerConfig
from lerobot.common.optim.schedulers import LRSchedulerConfig
from lerobot.common.utils.hub import HubMixin
from lerobot.common.utils.utils import auto_select_torch_device, is_amp_available
from lerobot.configs import parser
from lerobot.configs.default import DatasetConfig, EvalConfig, WandBConfig
from lerobot.configs.policies import PreTrainedConfig
@@ -48,10 +46,6 @@ class TrainPipelineConfig(HubMixin):
# Note that when resuming a run, the default behavior is to use the configuration from the checkpoint,
# regardless of what's provided with the training command at the time of resumption.
resume: bool = False
device: str | None = None # cuda | cpu | mp
# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
# automatic gradient scaling is used.
use_amp: bool = False
# `seed` is used for training (eg: model initialization, dataset shuffling)
# AND for the evaluation environments.
seed: int | None = 1000
@@ -74,18 +68,6 @@ class TrainPipelineConfig(HubMixin):
self.checkpoint_path = None
def validate(self):
if not self.device:
logging.warning("No device specified, trying to infer device automatically")
device = auto_select_torch_device()
self.device = device.type
# Automatically deactivate AMP if necessary
if self.use_amp and not is_amp_available(self.device):
logging.warning(
f"Automatic Mixed Precision (amp) is not available on device '{self.device}'. Deactivating AMP."
)
self.use_amp = False
# HACK: We parse again the cli args here to get the pretrained paths if there was some.
policy_path = parser.get_path_arg("policy")
if policy_path: