镜像自地址
https://github.com/MINT-SJTU/Evo-RL.git
已同步 2026-11-20 18:17:06 +00:00
snapshot: sync mint main from current workspace 2026-03-03 16:32:56 +0800
这个提交包含在:
@ -0,0 +1,88 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
from types import SimpleNamespace
|
||||
|
||||
from lerobot.rl.acp_dataset_stats import compute_acp_indicator_stats
|
||||
|
||||
|
||||
class DummyHFDataset:
|
||||
def __init__(self, columns: dict[str, list]):
|
||||
self._columns = columns
|
||||
self.column_names = list(columns.keys())
|
||||
|
||||
def __getitem__(self, key: str):
|
||||
return self._columns[key]
|
||||
|
||||
|
||||
def test_compute_acp_indicator_stats_from_meta_stats():
|
||||
dataset = SimpleNamespace(
|
||||
meta=SimpleNamespace(
|
||||
stats={
|
||||
"complementary_info.acp_indicator": {
|
||||
"mean": [0.3],
|
||||
"count": [10],
|
||||
}
|
||||
}
|
||||
)
|
||||
)
|
||||
|
||||
stats = compute_acp_indicator_stats(dataset, "complementary_info.acp_indicator")
|
||||
|
||||
assert stats is not None
|
||||
assert stats.source == "meta.stats"
|
||||
assert stats.positive_ratio == 0.3
|
||||
assert stats.total_count == 10
|
||||
assert stats.positive_count == 3
|
||||
assert stats.invalid_count == -1
|
||||
|
||||
|
||||
def test_compute_acp_indicator_stats_from_dataset_scan():
|
||||
dataset = SimpleNamespace(
|
||||
meta=SimpleNamespace(stats={}),
|
||||
hf_dataset=DummyHFDataset(
|
||||
{
|
||||
"complementary_info.acp_indicator": [[1], [0], [1], [2]],
|
||||
}
|
||||
),
|
||||
)
|
||||
|
||||
stats = compute_acp_indicator_stats(dataset, "complementary_info.acp_indicator")
|
||||
|
||||
assert stats is not None
|
||||
assert stats.source == "hf_dataset_scan"
|
||||
assert stats.total_count == 4
|
||||
assert stats.positive_count == 2
|
||||
assert stats.positive_ratio == 0.5
|
||||
assert stats.invalid_count == 1
|
||||
|
||||
|
||||
def test_compute_acp_indicator_stats_returns_none_when_missing_field():
|
||||
dataset = SimpleNamespace(
|
||||
meta=SimpleNamespace(stats={}),
|
||||
hf_dataset=DummyHFDataset({"task": ["pick", "place"]}),
|
||||
)
|
||||
|
||||
stats = compute_acp_indicator_stats(dataset, "complementary_info.acp_indicator")
|
||||
assert stats is None
|
||||
|
||||
|
||||
def test_compute_acp_indicator_stats_prefers_meta_stats():
|
||||
dataset = SimpleNamespace(
|
||||
meta=SimpleNamespace(
|
||||
stats={
|
||||
"complementary_info.acp_indicator": {
|
||||
"mean": [0.75],
|
||||
"count": [8],
|
||||
}
|
||||
}
|
||||
),
|
||||
hf_dataset=DummyHFDataset({"complementary_info.acp_indicator": [0, 0, 0, 0]}),
|
||||
)
|
||||
|
||||
stats = compute_acp_indicator_stats(dataset, "complementary_info.acp_indicator")
|
||||
|
||||
assert stats is not None
|
||||
assert stats.source == "meta.stats"
|
||||
assert stats.positive_ratio == 0.75
|
||||
assert stats.total_count == 8
|
||||
assert stats.positive_count == 6
|
||||
@ -0,0 +1,132 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
from unittest.mock import patch
|
||||
|
||||
import torch
|
||||
|
||||
from lerobot.configs.train import ACPConfig
|
||||
from lerobot.configs.types import FeatureType, PolicyFeature
|
||||
from lerobot.policies.pi05.configuration_pi05 import PI05Config
|
||||
from lerobot.policies.pi05.processor_pi05 import make_pi05_pre_post_processors
|
||||
from lerobot.rl.acp_hook import build_acp_raw_batch_hook
|
||||
from lerobot.utils.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS
|
||||
from tests.utils import require_package
|
||||
|
||||
|
||||
class TrackingTokenizer:
|
||||
def __init__(self):
|
||||
self.calls: list[list[str]] = []
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
text,
|
||||
max_length=512,
|
||||
truncation=True,
|
||||
padding="max_length",
|
||||
padding_side="right",
|
||||
return_tensors="pt",
|
||||
**kwargs,
|
||||
):
|
||||
del truncation, padding, padding_side, return_tensors, kwargs
|
||||
texts = [text] if isinstance(text, str) else list(text)
|
||||
self.calls.append(texts)
|
||||
|
||||
batch_size = len(texts)
|
||||
input_ids = torch.zeros(batch_size, max_length, dtype=torch.long)
|
||||
attention_mask = torch.zeros(batch_size, max_length, dtype=torch.long)
|
||||
for i, t in enumerate(texts):
|
||||
seq_len = min(max_length, max(1, len(t.split())))
|
||||
input_ids[i, :seq_len] = torch.arange(1, seq_len + 1, dtype=torch.long)
|
||||
attention_mask[i, :seq_len] = 1
|
||||
|
||||
if isinstance(text, str):
|
||||
return {
|
||||
"input_ids": input_ids[0],
|
||||
"attention_mask": attention_mask[0],
|
||||
}
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
}
|
||||
|
||||
|
||||
def _make_pi05_preprocessor():
|
||||
cfg = PI05Config(
|
||||
max_state_dim=4,
|
||||
max_action_dim=2,
|
||||
dtype="float32",
|
||||
device="cpu",
|
||||
)
|
||||
cfg.input_features = {
|
||||
"observation.state": PolicyFeature(type=FeatureType.STATE, shape=(4,)),
|
||||
}
|
||||
cfg.output_features = {
|
||||
"action": PolicyFeature(type=FeatureType.ACTION, shape=(2,)),
|
||||
}
|
||||
preprocessor, _ = make_pi05_pre_post_processors(config=cfg, dataset_stats=None)
|
||||
return preprocessor
|
||||
|
||||
|
||||
@require_package("transformers")
|
||||
@patch("lerobot.processor.tokenizer_processor.AutoTokenizer")
|
||||
def test_acp_prompt_reaches_pi05_tokenizer(mock_auto_tokenizer):
|
||||
tracking_tokenizer = TrackingTokenizer()
|
||||
mock_auto_tokenizer.from_pretrained.return_value = tracking_tokenizer
|
||||
|
||||
preprocessor = _make_pi05_preprocessor()
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(
|
||||
enable=True,
|
||||
indicator_field="complementary_info.acp_indicator",
|
||||
positive_tag="Advantage: positive",
|
||||
negative_tag="Advantage: negative",
|
||||
indicator_dropout_prob=0.0,
|
||||
),
|
||||
seed=123,
|
||||
)
|
||||
batch = {
|
||||
"observation.state": torch.tensor(
|
||||
[
|
||||
[0.1, -0.2, 0.0, 0.3],
|
||||
[-0.4, 0.5, -0.6, 0.7],
|
||||
],
|
||||
dtype=torch.float32,
|
||||
),
|
||||
"task": ["Pick bottle", "Place bottle"],
|
||||
"complementary_info.acp_indicator": torch.tensor([1, 0], dtype=torch.int64),
|
||||
}
|
||||
|
||||
conditioned_batch = hook(batch, 0)
|
||||
out = preprocessor(conditioned_batch)
|
||||
|
||||
prompts = tracking_tokenizer.calls[-1]
|
||||
assert len(prompts) == 2
|
||||
assert "Advantage: positive" in prompts[0]
|
||||
assert "Advantage: negative" in prompts[1]
|
||||
assert prompts[0].startswith("Task: Pick bottle")
|
||||
assert "State:" in prompts[0]
|
||||
assert OBS_LANGUAGE_TOKENS in out
|
||||
assert OBS_LANGUAGE_ATTENTION_MASK in out
|
||||
assert out[OBS_LANGUAGE_TOKENS].shape[0] == 2
|
||||
|
||||
|
||||
@require_package("transformers")
|
||||
@patch("lerobot.processor.tokenizer_processor.AutoTokenizer")
|
||||
def test_without_acp_no_advantage_tag_in_pi05_tokenizer(mock_auto_tokenizer):
|
||||
tracking_tokenizer = TrackingTokenizer()
|
||||
mock_auto_tokenizer.from_pretrained.return_value = tracking_tokenizer
|
||||
|
||||
preprocessor = _make_pi05_preprocessor()
|
||||
batch = {
|
||||
"observation.state": torch.tensor(
|
||||
[[0.0, 0.1, -0.1, 0.2]],
|
||||
dtype=torch.float32,
|
||||
),
|
||||
"task": ["Pick bottle"],
|
||||
}
|
||||
_ = preprocessor(batch)
|
||||
|
||||
prompts = tracking_tokenizer.calls[-1]
|
||||
assert len(prompts) == 1
|
||||
assert "Advantage:" not in prompts[0]
|
||||
assert prompts[0].startswith("Task: Pick bottle")
|
||||
@ -0,0 +1,157 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
from torch.utils.data import DataLoader
|
||||
|
||||
from lerobot.configs.train import ACPConfig
|
||||
from lerobot.datasets.lerobot_dataset import LeRobotDataset
|
||||
from lerobot.rl.acp_hook import build_acp_raw_batch_hook
|
||||
|
||||
|
||||
def test_acp_hook_accepts_integer_indicators():
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(
|
||||
enable=True,
|
||||
indicator_field="acp_indicator",
|
||||
positive_tag="POS",
|
||||
negative_tag="NEG",
|
||||
indicator_dropout_prob=0.0,
|
||||
),
|
||||
seed=42,
|
||||
)
|
||||
batch = {
|
||||
"task": ["pick bottle", "place bottle"],
|
||||
"acp_indicator": torch.tensor([1, 0], dtype=torch.int64),
|
||||
}
|
||||
|
||||
out = hook(batch, 0)
|
||||
assert out["task"] == [
|
||||
"pick bottle\nPOS",
|
||||
"place bottle\nNEG",
|
||||
]
|
||||
|
||||
|
||||
def test_acp_hook_injects_tags():
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(
|
||||
enable=True,
|
||||
indicator_field="complementary_info.acp_indicator",
|
||||
positive_tag="POS",
|
||||
negative_tag="NEG",
|
||||
indicator_dropout_prob=0.0,
|
||||
),
|
||||
seed=42,
|
||||
)
|
||||
batch = {
|
||||
"task": ["pick bottle", "place bottle"],
|
||||
"complementary_info.acp_indicator": torch.tensor([1, 0], dtype=torch.int64),
|
||||
}
|
||||
|
||||
out = hook(batch, 0)
|
||||
assert out["task"] == ["pick bottle\nPOS", "place bottle\nNEG"]
|
||||
|
||||
|
||||
def test_acp_hook_dropout_keeps_original_task():
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(
|
||||
enable=True,
|
||||
indicator_field="complementary_info.acp_indicator",
|
||||
positive_tag="POS",
|
||||
negative_tag="NEG",
|
||||
indicator_dropout_prob=1.0,
|
||||
),
|
||||
seed=42,
|
||||
)
|
||||
batch = {
|
||||
"task": ["pick bottle", "place bottle"],
|
||||
"complementary_info.acp_indicator": torch.tensor([1, 0], dtype=torch.int64),
|
||||
}
|
||||
|
||||
out = hook(batch, 0)
|
||||
assert out["task"] == ["pick bottle", "place bottle"]
|
||||
|
||||
|
||||
def test_acp_hook_missing_indicator_skips():
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(enable=True, indicator_field="missing_field"),
|
||||
seed=42,
|
||||
)
|
||||
batch = {"task": ["pick bottle"]}
|
||||
with pytest.raises(KeyError, match="missing_field"):
|
||||
hook(batch, 0)
|
||||
|
||||
|
||||
def test_acp_hook_non_integer_type_raises():
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(
|
||||
enable=True,
|
||||
indicator_field="acp_indicator",
|
||||
positive_tag="POS",
|
||||
negative_tag="NEG",
|
||||
indicator_dropout_prob=0.0,
|
||||
),
|
||||
seed=42,
|
||||
)
|
||||
batch = {
|
||||
"task": ["pick bottle", "place bottle"],
|
||||
"acp_indicator": torch.tensor([1.0, 0.0], dtype=torch.float32),
|
||||
}
|
||||
with pytest.raises(TypeError, match="integer 0/1"):
|
||||
hook(batch, 0)
|
||||
|
||||
|
||||
def test_acp_hook_non_binary_value_raises():
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(
|
||||
enable=True,
|
||||
indicator_field="acp_indicator",
|
||||
positive_tag="POS",
|
||||
negative_tag="NEG",
|
||||
indicator_dropout_prob=0.0,
|
||||
),
|
||||
seed=42,
|
||||
)
|
||||
batch = {
|
||||
"task": ["pick bottle", "place bottle"],
|
||||
"acp_indicator": torch.tensor([2, 0], dtype=torch.int64),
|
||||
}
|
||||
with pytest.raises(ValueError, match="must be 0 or 1"):
|
||||
hook(batch, 0)
|
||||
|
||||
|
||||
def _load_real_batch_or_skip(batch_size: int) -> dict:
|
||||
repo_id = "local/eval_so101_pick_place_bottle_v0"
|
||||
try:
|
||||
dataset = LeRobotDataset(repo_id, download_videos=False)
|
||||
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=0)
|
||||
return next(iter(dataloader))
|
||||
except Exception as e:
|
||||
pytest.skip(f"Cannot load local dataset '{repo_id}': {e}")
|
||||
|
||||
|
||||
@pytest.mark.parametrize("batch_size", [1, 3])
|
||||
def test_acp_hook_with_real_local_dataset_batch(batch_size: int):
|
||||
batch = _load_real_batch_or_skip(batch_size)
|
||||
assert "task" in batch
|
||||
assert "complementary_info.acp_indicator" not in batch
|
||||
|
||||
indicators = torch.randint(0, 2, (batch_size,), dtype=torch.int64)
|
||||
batch["complementary_info.acp_indicator"] = indicators
|
||||
|
||||
hook = build_acp_raw_batch_hook(
|
||||
ACPConfig(
|
||||
enable=True,
|
||||
indicator_field="complementary_info.acp_indicator",
|
||||
positive_tag="POS",
|
||||
negative_tag="NEG",
|
||||
indicator_dropout_prob=0.0,
|
||||
),
|
||||
seed=42,
|
||||
)
|
||||
out = hook(batch, 0)
|
||||
|
||||
assert len(out["task"]) == batch_size
|
||||
for i in range(batch_size):
|
||||
expected_tag = "POS" if int(indicators[i].item()) == 1 else "NEG"
|
||||
assert out["task"][i].endswith(expected_tag)
|
||||
211
tests/training/test_multi_gpu.py
普通文件
211
tests/training/test_multi_gpu.py
普通文件
@ -0,0 +1,211 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
"""
|
||||
Multi-GPU Training Tests
|
||||
|
||||
This module tests multi-GPU training functionality with accelerate.
|
||||
These tests are designed to run on machines with 2+ GPUs and are executed
|
||||
in the nightly CI workflow.
|
||||
|
||||
The tests automatically generate accelerate configs and launch training
|
||||
with subprocess to properly test the distributed training environment.
|
||||
"""
|
||||
|
||||
import os
|
||||
import subprocess
|
||||
import tempfile
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from lerobot.datasets.lerobot_dataset import LeRobotDataset
|
||||
|
||||
|
||||
def get_num_available_gpus():
|
||||
"""Returns the number of available GPUs."""
|
||||
if not torch.cuda.is_available():
|
||||
return 0
|
||||
return torch.cuda.device_count()
|
||||
|
||||
|
||||
def download_dataset(repo_id, episodes):
|
||||
"""
|
||||
Pre-download dataset to avoid race conditions in multi-GPU training.
|
||||
|
||||
Args:
|
||||
repo_id: HuggingFace dataset repository ID
|
||||
episodes: List of episode indices to download
|
||||
"""
|
||||
# Simply instantiating the dataset will download it
|
||||
_ = LeRobotDataset(repo_id, episodes=episodes)
|
||||
print(f"Dataset {repo_id} downloaded successfully")
|
||||
|
||||
|
||||
def run_accelerate_training(config_args, num_processes=4, temp_dir=None):
|
||||
"""
|
||||
Helper function to run training with accelerate launch.
|
||||
|
||||
Args:
|
||||
config_args: List of config arguments to pass to lerobot_train.py
|
||||
num_processes: Number of processes (GPUs) to use
|
||||
temp_dir: Temporary directory for outputs
|
||||
|
||||
Returns:
|
||||
subprocess.CompletedProcess result
|
||||
"""
|
||||
|
||||
config_path = Path(temp_dir) / "accelerate_config.yaml"
|
||||
|
||||
# Write YAML config
|
||||
with open(config_path, "w") as f:
|
||||
f.write("compute_environment: LOCAL_MACHINE\n")
|
||||
f.write("distributed_type: MULTI_GPU\n")
|
||||
f.write("mixed_precision: 'no'\n")
|
||||
f.write(f"num_processes: {num_processes}\n")
|
||||
f.write("use_cpu: false\n")
|
||||
f.write("gpu_ids: all\n")
|
||||
f.write("downcast_bf16: 'no'\n")
|
||||
f.write("machine_rank: 0\n")
|
||||
f.write("main_training_function: main\n")
|
||||
f.write("num_machines: 1\n")
|
||||
f.write("rdzv_backend: static\n")
|
||||
f.write("same_network: true\n")
|
||||
|
||||
cmd = [
|
||||
"accelerate",
|
||||
"launch",
|
||||
"--config_file",
|
||||
str(config_path),
|
||||
"-m",
|
||||
"lerobot.scripts.lerobot_train",
|
||||
] + config_args
|
||||
|
||||
result = subprocess.run(
|
||||
cmd,
|
||||
capture_output=True,
|
||||
text=True,
|
||||
env={**os.environ, "CUDA_VISIBLE_DEVICES": ",".join(map(str, range(num_processes)))},
|
||||
)
|
||||
|
||||
return result
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
get_num_available_gpus() < 2,
|
||||
reason="Multi-GPU tests require at least 2 GPUs",
|
||||
)
|
||||
class TestMultiGPUTraining:
|
||||
"""Test suite for multi-GPU training functionality."""
|
||||
|
||||
def test_basic_multi_gpu_training(self):
|
||||
"""
|
||||
Test that basic multi-GPU training runs successfully.
|
||||
Verifies that the training completes without errors.
|
||||
"""
|
||||
# Pre-download dataset to avoid race conditions
|
||||
download_dataset("lerobot/pusht", episodes=[0])
|
||||
|
||||
with tempfile.TemporaryDirectory() as temp_dir:
|
||||
output_dir = Path(temp_dir) / "outputs"
|
||||
|
||||
config_args = [
|
||||
"--dataset.repo_id=lerobot/pusht",
|
||||
"--dataset.episodes=[0]",
|
||||
"--policy.type=act",
|
||||
"--policy.device=cuda",
|
||||
"--policy.push_to_hub=false",
|
||||
f"--output_dir={output_dir}",
|
||||
"--batch_size=4",
|
||||
"--steps=10",
|
||||
"--eval_freq=-1",
|
||||
"--log_freq=5",
|
||||
"--save_freq=10",
|
||||
"--seed=42",
|
||||
"--num_workers=0",
|
||||
]
|
||||
|
||||
result = run_accelerate_training(config_args, num_processes=4, temp_dir=temp_dir)
|
||||
|
||||
# Check that training completed successfully
|
||||
assert result.returncode == 0, (
|
||||
f"Multi-GPU training failed with return code {result.returncode}\n"
|
||||
f"STDOUT:\n{result.stdout}\n"
|
||||
f"STDERR:\n{result.stderr}"
|
||||
)
|
||||
|
||||
# Verify checkpoint was saved
|
||||
checkpoints_dir = output_dir / "checkpoints"
|
||||
assert checkpoints_dir.exists(), "Checkpoints directory was not created"
|
||||
|
||||
# Verify that training completed
|
||||
assert "End of training" in result.stdout or "End of training" in result.stderr
|
||||
|
||||
def test_checkpoint_saving_multi_gpu(self):
|
||||
"""
|
||||
Test that checkpoints are correctly saved during multi-GPU training.
|
||||
Only the main process (rank 0) should save checkpoints.
|
||||
"""
|
||||
# Pre-download dataset to avoid race conditions
|
||||
download_dataset("lerobot/pusht", episodes=[0])
|
||||
|
||||
with tempfile.TemporaryDirectory() as temp_dir:
|
||||
output_dir = Path(temp_dir) / "outputs"
|
||||
|
||||
config_args = [
|
||||
"--dataset.repo_id=lerobot/pusht",
|
||||
"--dataset.episodes=[0]",
|
||||
"--policy.type=act",
|
||||
"--policy.device=cuda",
|
||||
"--policy.push_to_hub=false",
|
||||
f"--output_dir={output_dir}",
|
||||
"--batch_size=4",
|
||||
"--steps=20",
|
||||
"--eval_freq=-1",
|
||||
"--log_freq=5",
|
||||
"--save_freq=10",
|
||||
"--seed=42",
|
||||
"--num_workers=0",
|
||||
]
|
||||
|
||||
result = run_accelerate_training(config_args, num_processes=2, temp_dir=temp_dir)
|
||||
|
||||
assert result.returncode == 0, (
|
||||
f"Training failed:\nSTDOUT:\n{result.stdout}\n\nSTDERR:\n{result.stderr}"
|
||||
)
|
||||
|
||||
# Verify checkpoint directory exists
|
||||
checkpoints_dir = output_dir / "checkpoints"
|
||||
assert checkpoints_dir.exists(), "Checkpoints directory not created"
|
||||
|
||||
# Count checkpoint directories (should have checkpoint at step 10 and 20)
|
||||
checkpoint_dirs = [d for d in checkpoints_dir.iterdir() if d.is_dir()]
|
||||
assert len(checkpoint_dirs) >= 1, f"Expected at least 1 checkpoint, found {len(checkpoint_dirs)}"
|
||||
|
||||
# Verify checkpoint contents
|
||||
for checkpoint_dir in checkpoint_dirs:
|
||||
# Check for model files
|
||||
model_files = list(checkpoint_dir.rglob("*.safetensors"))
|
||||
assert len(model_files) > 0, f"No model files in checkpoint {checkpoint_dir}"
|
||||
|
||||
# Check for training state
|
||||
training_state_dir = checkpoint_dir / "training_state"
|
||||
assert training_state_dir.exists(), f"No training state in checkpoint {checkpoint_dir}"
|
||||
|
||||
# Verify optimizer state exists
|
||||
optimizer_state = training_state_dir / "optimizer_state.safetensors"
|
||||
assert optimizer_state.exists(), f"No optimizer state in checkpoint {checkpoint_dir}"
|
||||
@ -0,0 +1,157 @@
|
||||
#!/usr/bin/env python
|
||||
|
||||
# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
"""
|
||||
Visual Feature Consistency Tests
|
||||
|
||||
This module tests the `validate_visual_features_consistency` function,
|
||||
which ensures that visual features (camera observations) in a dataset/env
|
||||
match the expectations defined in a policy configuration.
|
||||
|
||||
The purpose of this check is to prevent mismatches between what a policy expects
|
||||
(e.g., `observation.images.camera1`, `camera2`, `camera3`) and what a dataset or
|
||||
environment actually provides (e.g., `observation.images.top`, `side`, or fewer cameras).
|
||||
"""
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
import numpy as np
|
||||
import pytest
|
||||
|
||||
from lerobot.configs.default import DatasetConfig
|
||||
from lerobot.configs.policies import PreTrainedConfig
|
||||
from lerobot.configs.train import TrainPipelineConfig
|
||||
from lerobot.datasets.lerobot_dataset import LeRobotDataset
|
||||
from lerobot.policies.factory import make_policy_config
|
||||
from lerobot.scripts.lerobot_train import train
|
||||
from lerobot.utils.utils import auto_select_torch_device
|
||||
|
||||
pytest.importorskip("transformers")
|
||||
|
||||
DUMMY_REPO_ID = "dummy/repo"
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def temp_dir(tmp_path):
|
||||
return tmp_path
|
||||
|
||||
|
||||
DUMMY_STATE_DIM = 6
|
||||
DUMMY_ACTION_DIM = 6
|
||||
IMAGE_SIZE = 8
|
||||
DEVICE = auto_select_torch_device()
|
||||
|
||||
|
||||
def make_dummy_dataset(camera_keys, tmp_path):
|
||||
"""Creates a minimal dummy dataset for testing rename_mapping logic."""
|
||||
features = {
|
||||
"action": {"dtype": "float32", "shape": (DUMMY_ACTION_DIM,), "names": None},
|
||||
"observation.state": {"dtype": "float32", "shape": (DUMMY_STATE_DIM,), "names": None},
|
||||
}
|
||||
for cam in camera_keys:
|
||||
features[f"observation.images.{cam}"] = {
|
||||
"dtype": "image",
|
||||
"shape": (IMAGE_SIZE, IMAGE_SIZE, 3),
|
||||
"names": ["height", "width", "channel"],
|
||||
}
|
||||
dataset = LeRobotDataset.create(
|
||||
repo_id=DUMMY_REPO_ID,
|
||||
fps=30,
|
||||
features=features,
|
||||
root=tmp_path / "_dataset",
|
||||
)
|
||||
root = tmp_path / "_dataset"
|
||||
for ep_idx in range(2):
|
||||
for _ in range(3):
|
||||
frame = {
|
||||
"action": np.random.randn(DUMMY_ACTION_DIM).astype(np.float32),
|
||||
"observation.state": np.random.randn(DUMMY_STATE_DIM).astype(np.float32),
|
||||
}
|
||||
for cam in camera_keys:
|
||||
frame[f"observation.images.{cam}"] = np.random.randint(
|
||||
0, 255, size=(IMAGE_SIZE, IMAGE_SIZE, 3), dtype=np.uint8
|
||||
)
|
||||
frame["task"] = f"task_{ep_idx}"
|
||||
dataset.add_frame(frame)
|
||||
dataset.save_episode()
|
||||
|
||||
dataset.finalize()
|
||||
return dataset, root
|
||||
|
||||
|
||||
def custom_validate(train_config: TrainPipelineConfig, policy_path: str, empty_cameras: int):
|
||||
train_config.policy = PreTrainedConfig.from_pretrained(policy_path)
|
||||
train_config.policy.pretrained_path = Path(policy_path)
|
||||
# override empty_cameras and push_to_hub for testing
|
||||
train_config.policy.empty_cameras = empty_cameras
|
||||
train_config.policy.push_to_hub = False
|
||||
if train_config.use_policy_training_preset:
|
||||
train_config.optimizer = train_config.policy.get_optimizer_preset()
|
||||
train_config.scheduler = train_config.policy.get_scheduler_preset()
|
||||
return train_config
|
||||
|
||||
|
||||
@pytest.mark.skip(reason="Skipping this test as it results OOM")
|
||||
@pytest.mark.parametrize(
|
||||
"camera_keys, empty_cameras, rename_map, expect_success",
|
||||
[
|
||||
# case 1: dataset has fewer cameras than policy (3 instead of 4), but we specify empty_cameras=1 for smolvla, pi0, pi05
|
||||
(["camera1", "camera2", "camera3"], 1, {}, True),
|
||||
# case 2: dataset has 2 cameras with different names, rename_mapping provided
|
||||
(
|
||||
["top", "side"],
|
||||
0,
|
||||
{
|
||||
"observation.images.top": "observation.images.camera1",
|
||||
"observation.images.side": "observation.images.camera2",
|
||||
},
|
||||
True,
|
||||
),
|
||||
# case 3: dataset has 2 cameras, policy expects 3, names do not match, no empty_cameras
|
||||
(["top", "side"], 0, {}, False),
|
||||
# TODO: case 4: dataset has 2 cameras, policy expects 3, no rename_map, no empty_cameras, should raise for smolvla
|
||||
# (["camera1", "camera2"], 0, {}, False),
|
||||
],
|
||||
)
|
||||
def test_train_with_camera_mismatch(camera_keys, empty_cameras, rename_map, expect_success, tmp_path):
|
||||
"""Tests that training works or fails depending on camera/feature alignment."""
|
||||
|
||||
_dataset, root = make_dummy_dataset(camera_keys, tmp_path)
|
||||
pretrained_path = "lerobot/smolvla_base"
|
||||
dataset_config = DatasetConfig(repo_id=DUMMY_REPO_ID, root=root)
|
||||
policy_config = make_policy_config(
|
||||
"smolvla",
|
||||
optimizer_lr=0.01,
|
||||
push_to_hub=False,
|
||||
pretrained_path=pretrained_path,
|
||||
device=DEVICE,
|
||||
)
|
||||
policy_config.empty_cameras = empty_cameras
|
||||
train_config = TrainPipelineConfig(
|
||||
dataset=dataset_config,
|
||||
policy=policy_config,
|
||||
rename_map=rename_map,
|
||||
output_dir=tmp_path / "_output",
|
||||
steps=1,
|
||||
)
|
||||
train_config = custom_validate(train_config, policy_path=pretrained_path, empty_cameras=empty_cameras)
|
||||
# HACK: disable the internal CLI validation step for tests, we did it with custom_validate
|
||||
train_config.validate = lambda: None
|
||||
if expect_success:
|
||||
train(train_config)
|
||||
else:
|
||||
with pytest.raises(ValueError):
|
||||
train(train_config)
|
||||
在新工单中引用
屏蔽一个用户