mirror of
https://github.com/huggingface/lerobot.git
synced 2026-05-20 02:59:50 +00:00
Remove previous pi0 and rename pi0_openpi and pi05_openpi
This commit is contained in:
Pepijn
@@ -13,7 +13,7 @@ pytestmark = pytest.mark.skipif(
|
||||
reason="This test requires local OpenPI installation and is not meant for CI",
|
||||
)
|
||||
|
||||
from lerobot.policies.pi05_openpi import PI05OpenPIConfig, PI05OpenPIPolicy # noqa: E402
|
||||
from lerobot.policies.pi05 import PI05Config, PI05Policy # noqa: E402
|
||||
from tests.utils import require_cuda # noqa: E402
|
||||
|
||||
|
||||
@@ -22,7 +22,7 @@ def test_pi05_model_architecture():
|
||||
"""Test that pi05=True creates the correct model architecture."""
|
||||
|
||||
# Create config
|
||||
config = PI05OpenPIConfig(
|
||||
config = PI05Config(
|
||||
max_action_dim=7,
|
||||
max_state_dim=14,
|
||||
dtype="float32",
|
||||
@@ -73,7 +73,7 @@ def test_pi05_model_architecture():
|
||||
}
|
||||
|
||||
# Instantiate policy
|
||||
policy = PI05OpenPIPolicy(config, dataset_stats)
|
||||
policy = PI05Policy(config, dataset_stats)
|
||||
|
||||
# Verify pi05 model components exist
|
||||
# Check that time_mlp layers exist (for AdaRMS conditioning)
|
||||
@@ -104,7 +104,7 @@ def test_pi05_forward_pass():
|
||||
"""Test forward pass with"""
|
||||
|
||||
# Create config
|
||||
config = PI05OpenPIConfig(
|
||||
config = PI05Config(
|
||||
max_action_dim=7,
|
||||
max_state_dim=14,
|
||||
dtype="float32",
|
||||
@@ -150,7 +150,7 @@ def test_pi05_forward_pass():
|
||||
}
|
||||
|
||||
# Instantiate policy
|
||||
policy = PI05OpenPIPolicy(config, dataset_stats)
|
||||
policy = PI05Policy(config, dataset_stats)
|
||||
|
||||
# Create test batch
|
||||
batch_size = 2
|
||||
|
||||
@@ -14,7 +14,7 @@ pytestmark = pytest.mark.skipif(
|
||||
)
|
||||
|
||||
from lerobot.policies.factory import make_policy_config # noqa: E402
|
||||
from lerobot.policies.pi0_openpi import PI0OpenPIConfig, PI0OpenPIPolicy # noqa: E402
|
||||
from lerobot.policies.pi0 import PI0OpenPIConfig, PI0OpenPIPolicy # noqa: E402
|
||||
from tests.utils import require_cuda # noqa: E402
|
||||
|
||||
|
||||
@@ -96,7 +96,7 @@ def test_config_creation():
|
||||
"""Test policy config creation through factory."""
|
||||
try:
|
||||
config = make_policy_config(
|
||||
policy_type="pi0_openpi",
|
||||
policy_type="pi0",
|
||||
max_action_dim=7,
|
||||
max_state_dim=14,
|
||||
)
|
||||
|
||||
@@ -21,7 +21,7 @@ from openpi.models_pytorch import preprocessing_pytorch as openpi_preprocessing
|
||||
from openpi.models_pytorch.pi0_pytorch import PI0Pytorch # noqa: E402
|
||||
from transformers import AutoTokenizer # noqa: E402
|
||||
|
||||
from lerobot.policies.pi0_openpi import PI0OpenPIConfig, PI0OpenPIPolicy # noqa: E402
|
||||
from lerobot.policies.pi0 import PI0Config, PI0Policy # noqa: E402
|
||||
|
||||
DUMMY_ACTION_DIM = 32
|
||||
DUMMY_STATE_DIM = 32
|
||||
@@ -68,9 +68,7 @@ class PI0BaseOriginalConfig:
|
||||
def instantiate_lerobot_pi0(from_pretrained: bool = False):
|
||||
if from_pretrained:
|
||||
# Load the policy first
|
||||
policy = PI0OpenPIPolicy.from_pretrained(
|
||||
pretrained_name_or_path="pepijn223/pi0_base_fp32", strict=True
|
||||
)
|
||||
policy = PI0Policy.from_pretrained(pretrained_name_or_path="pepijn223/pi0_base_fp32", strict=True)
|
||||
# Then reinitialize the normalization with proper stats
|
||||
from lerobot.policies.normalize import Normalize, Unnormalize
|
||||
|
||||
@@ -84,10 +82,8 @@ def instantiate_lerobot_pi0(from_pretrained: bool = False):
|
||||
policy.config.output_features, policy.config.normalization_mapping, DUMMY_DATASET_STATS
|
||||
)
|
||||
else:
|
||||
config = PI0OpenPIConfig(
|
||||
max_action_dim=DUMMY_ACTION_DIM, max_state_dim=DUMMY_STATE_DIM, dtype="float32"
|
||||
)
|
||||
policy = PI0OpenPIPolicy(config, DUMMY_DATASET_STATS)
|
||||
config = PI0Config(max_action_dim=DUMMY_ACTION_DIM, max_state_dim=DUMMY_STATE_DIM, dtype="float32")
|
||||
policy = PI0Policy(config, DUMMY_DATASET_STATS)
|
||||
policy.to(DEVICE)
|
||||
return policy
|
||||
|
||||
|
||||
@@ -18,8 +18,8 @@ pytestmark = pytest.mark.skipif(
|
||||
reason="This test requires HuggingFace authentication and is not meant for CI",
|
||||
)
|
||||
|
||||
from lerobot.policies.pi0_openpi import PI0OpenPIPolicy # noqa: E402
|
||||
from lerobot.policies.pi05_openpi.modeling_pi05openpi import PI05OpenPIPolicy # noqa: E402
|
||||
from lerobot.policies.pi0 import PI0Policy # noqa: E402
|
||||
from lerobot.policies.pi05.modeling_pi05openpi import PI05Policy # noqa: E402
|
||||
|
||||
|
||||
def create_dummy_stats(config):
|
||||
@@ -48,13 +48,13 @@ def create_dummy_stats(config):
|
||||
# Test data for all 6 base models
|
||||
MODEL_TEST_PARAMS = [
|
||||
# PI0 models
|
||||
("pepijn223/pi0_base_fp32", "PI0", PI0OpenPIPolicy),
|
||||
("pepijn223/pi0_droid_fp32", "PI0", PI0OpenPIPolicy),
|
||||
("pepijn223/pi0_libero_fp32", "PI0", PI0OpenPIPolicy),
|
||||
("pepijn223/pi0_base_fp32", "PI0", PI0Policy),
|
||||
("pepijn223/pi0_droid_fp32", "PI0", PI0Policy),
|
||||
("pepijn223/pi0_libero_fp32", "PI0", PI0Policy),
|
||||
# PI0.5 models
|
||||
("pepijn223/pi05_base_fp32", "PI0.5", PI05OpenPIPolicy),
|
||||
("pepijn223/pi05_droid_fp32", "PI0.5", PI05OpenPIPolicy),
|
||||
("pepijn223/pi05_libero_fp32", "PI0.5", PI05OpenPIPolicy),
|
||||
("pepijn223/pi05_base_fp32", "PI0.5", PI05Policy),
|
||||
("pepijn223/pi05_droid_fp32", "PI0.5", PI05Policy),
|
||||
("pepijn223/pi05_libero_fp32", "PI0.5", PI05Policy),
|
||||
]
|
||||
|
||||
|
||||
@@ -65,7 +65,7 @@ def test_all_base_models_hub_loading(model_id, model_type, policy_class):
|
||||
Args:
|
||||
model_id: HuggingFace model ID (e.g., "pepijn223/pi0_base_fp32")
|
||||
model_type: Model type ("PI0" or "PI0.5")
|
||||
policy_class: Policy class to use (PI0OpenPIPolicy or PI05OpenPIPolicy)
|
||||
policy_class: Policy class to use (PI0Policy or PI05Policy)
|
||||
"""
|
||||
print(f"\n{'=' * 80}")
|
||||
print(f"Testing {model_type} model: {model_id}")
|
||||
|
||||
@@ -1,424 +0,0 @@
|
||||
#!/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.
|
||||
"""Tests for PI0 policy processor."""
|
||||
|
||||
from unittest.mock import patch
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
|
||||
from lerobot.constants import ACTION, OBS_IMAGE, OBS_STATE
|
||||
from lerobot.policies.pi0.configuration_pi0 import PI0Config
|
||||
from lerobot.policies.pi0.processor_pi0 import Pi0NewLineProcessor, make_pi0_pre_post_processors
|
||||
from lerobot.processor import (
|
||||
AddBatchDimensionProcessorStep,
|
||||
DeviceProcessorStep,
|
||||
EnvTransition,
|
||||
NormalizerProcessorStep,
|
||||
ProcessorStep,
|
||||
RenameObservationsProcessorStep,
|
||||
TransitionKey,
|
||||
UnnormalizerProcessorStep,
|
||||
)
|
||||
from lerobot.processor.converters import create_transition, transition_to_batch
|
||||
|
||||
|
||||
class MockTokenizerProcessorStep(ProcessorStep):
|
||||
"""Mock tokenizer processor step for testing."""
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
# Accept any arguments to mimic the real TokenizerProcessorStep interface
|
||||
pass
|
||||
|
||||
def __call__(self, transition: EnvTransition) -> EnvTransition:
|
||||
# Pass through transition unchanged
|
||||
return transition
|
||||
|
||||
def transform_features(self, features):
|
||||
# Pass through features unchanged
|
||||
return features
|
||||
|
||||
|
||||
def create_default_config():
|
||||
"""Create a default PI0 configuration for testing."""
|
||||
config = PI0Config()
|
||||
config.input_features = {
|
||||
OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(10,)),
|
||||
OBS_IMAGE: PolicyFeature(type=FeatureType.VISUAL, shape=(3, 224, 224)),
|
||||
}
|
||||
config.output_features = {
|
||||
ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(6,)),
|
||||
}
|
||||
config.normalization_mapping = {
|
||||
FeatureType.STATE: NormalizationMode.MEAN_STD,
|
||||
FeatureType.VISUAL: NormalizationMode.IDENTITY,
|
||||
FeatureType.ACTION: NormalizationMode.MIN_MAX,
|
||||
}
|
||||
config.device = "cpu"
|
||||
config.tokenizer_max_length = 128
|
||||
return config
|
||||
|
||||
|
||||
def create_default_stats():
|
||||
"""Create default dataset statistics for testing."""
|
||||
return {
|
||||
OBS_STATE: {"mean": torch.zeros(10), "std": torch.ones(10)},
|
||||
OBS_IMAGE: {}, # No normalization for images
|
||||
ACTION: {"min": torch.full((6,), -1.0), "max": torch.ones(6)},
|
||||
}
|
||||
|
||||
|
||||
def test_make_pi0_processor_basic():
|
||||
"""Test basic creation of PI0 processor."""
|
||||
config = create_default_config()
|
||||
stats = create_default_stats()
|
||||
|
||||
with patch("lerobot.policies.pi0.processor_pi0.TokenizerProcessorStep", MockTokenizerProcessorStep):
|
||||
preprocessor, postprocessor = make_pi0_pre_post_processors(
|
||||
config,
|
||||
stats,
|
||||
)
|
||||
|
||||
# Check processor names
|
||||
assert preprocessor.name == "policy_preprocessor"
|
||||
assert postprocessor.name == "policy_postprocessor"
|
||||
|
||||
# Check steps in preprocessor
|
||||
assert len(preprocessor.steps) == 6
|
||||
assert isinstance(preprocessor.steps[0], RenameObservationsProcessorStep)
|
||||
assert isinstance(preprocessor.steps[1], AddBatchDimensionProcessorStep)
|
||||
assert isinstance(preprocessor.steps[2], Pi0NewLineProcessor)
|
||||
# Step 3 would be TokenizerProcessorStep but it's mocked
|
||||
assert isinstance(preprocessor.steps[4], DeviceProcessorStep)
|
||||
assert isinstance(preprocessor.steps[5], NormalizerProcessorStep)
|
||||
|
||||
# Check steps in postprocessor
|
||||
assert len(postprocessor.steps) == 2
|
||||
assert isinstance(postprocessor.steps[0], UnnormalizerProcessorStep)
|
||||
assert isinstance(postprocessor.steps[1], DeviceProcessorStep)
|
||||
|
||||
|
||||
def test_pi0_newline_processor_single_task():
|
||||
"""Test Pi0NewLineProcessor with single task string."""
|
||||
processor = Pi0NewLineProcessor()
|
||||
|
||||
# Test with task that doesn't have newline
|
||||
transition = create_transition(complementary_data={"task": "test task"})
|
||||
result = processor(transition)
|
||||
assert result[TransitionKey.COMPLEMENTARY_DATA]["task"] == "test task\n"
|
||||
|
||||
# Test with task that already has newline
|
||||
transition = create_transition(complementary_data={"task": "test task\n"})
|
||||
result = processor(transition)
|
||||
assert result[TransitionKey.COMPLEMENTARY_DATA]["task"] == "test task\n"
|
||||
|
||||
|
||||
def test_pi0_newline_processor_list_of_tasks():
|
||||
"""Test Pi0NewLineProcessor with list of task strings."""
|
||||
processor = Pi0NewLineProcessor()
|
||||
|
||||
# Test with list of tasks
|
||||
tasks = ["task1", "task2\n", "task3"]
|
||||
transition = create_transition(complementary_data={"task": tasks})
|
||||
result = processor(transition)
|
||||
expected = ["task1\n", "task2\n", "task3\n"]
|
||||
assert result[TransitionKey.COMPLEMENTARY_DATA]["task"] == expected
|
||||
|
||||
|
||||
def test_pi0_newline_processor_empty_transition():
|
||||
"""Test Pi0NewLineProcessor with empty transition."""
|
||||
processor = Pi0NewLineProcessor()
|
||||
|
||||
# Test with no complementary_data
|
||||
transition = create_transition()
|
||||
result = processor(transition)
|
||||
assert result == transition
|
||||
|
||||
# Test with complementary_data but no task
|
||||
transition = create_transition(complementary_data={"other": "data"})
|
||||
result = processor(transition)
|
||||
assert result == transition
|
||||
|
||||
# Test with None task
|
||||
transition = create_transition(complementary_data={"task": None})
|
||||
result = processor(transition)
|
||||
assert result == transition
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
|
||||
def test_pi0_processor_cuda():
|
||||
"""Test PI0 processor with CUDA device."""
|
||||
config = create_default_config()
|
||||
config.device = "cuda"
|
||||
stats = create_default_stats()
|
||||
|
||||
# Mock the tokenizer processor to act as pass-through
|
||||
class MockTokenizerProcessorStep(ProcessorStep):
|
||||
def __init__(self, *args, **kwargs):
|
||||
pass
|
||||
|
||||
def __call__(self, transition):
|
||||
return transition
|
||||
|
||||
def state_dict(self):
|
||||
return {}
|
||||
|
||||
def load_state_dict(self, state):
|
||||
pass
|
||||
|
||||
def reset(self):
|
||||
pass
|
||||
|
||||
def get_config(self):
|
||||
return {"tokenizer_name": "google/paligemma-3b-pt-224"}
|
||||
|
||||
def transform_features(self, features):
|
||||
return features
|
||||
|
||||
with patch("lerobot.policies.pi0.processor_pi0.TokenizerProcessorStep", MockTokenizerProcessorStep):
|
||||
preprocessor, postprocessor = make_pi0_pre_post_processors(
|
||||
config,
|
||||
stats,
|
||||
)
|
||||
|
||||
# Create CPU data
|
||||
observation = {
|
||||
OBS_STATE: torch.randn(10),
|
||||
OBS_IMAGE: torch.randn(3, 224, 224),
|
||||
}
|
||||
action = torch.randn(6)
|
||||
transition = create_transition(observation, action, complementary_data={"task": "test task"})
|
||||
batch = transition_to_batch(transition)
|
||||
|
||||
# Process through preprocessor
|
||||
processed = preprocessor(batch)
|
||||
|
||||
# Check that data is on CUDA
|
||||
assert processed[OBS_STATE].device.type == "cuda"
|
||||
assert processed[OBS_IMAGE].device.type == "cuda"
|
||||
assert processed[TransitionKey.ACTION.value].device.type == "cuda"
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
|
||||
def test_pi0_processor_accelerate_scenario():
|
||||
"""Test PI0 processor in simulated Accelerate scenario."""
|
||||
config = create_default_config()
|
||||
config.device = "cuda:0"
|
||||
stats = create_default_stats()
|
||||
|
||||
# Mock the tokenizer processor to act as pass-through
|
||||
class MockTokenizerProcessorStep(ProcessorStep):
|
||||
def __init__(self, *args, **kwargs):
|
||||
pass
|
||||
|
||||
def __call__(self, transition):
|
||||
return transition
|
||||
|
||||
def state_dict(self):
|
||||
return {}
|
||||
|
||||
def load_state_dict(self, state):
|
||||
pass
|
||||
|
||||
def reset(self):
|
||||
pass
|
||||
|
||||
def get_config(self):
|
||||
return {"tokenizer_name": "google/paligemma-3b-pt-224"}
|
||||
|
||||
def transform_features(self, features):
|
||||
return features
|
||||
|
||||
with patch("lerobot.policies.pi0.processor_pi0.TokenizerProcessorStep", MockTokenizerProcessorStep):
|
||||
preprocessor, postprocessor = make_pi0_pre_post_processors(
|
||||
config,
|
||||
stats,
|
||||
)
|
||||
|
||||
# Simulate Accelerate: data already on GPU and batched
|
||||
device = torch.device("cuda:0")
|
||||
observation = {
|
||||
OBS_STATE: torch.randn(1, 10).to(device),
|
||||
OBS_IMAGE: torch.randn(1, 3, 224, 224).to(device),
|
||||
}
|
||||
action = torch.randn(1, 6).to(device)
|
||||
transition = create_transition(observation, action, complementary_data={"task": ["test task"]})
|
||||
batch = transition_to_batch(transition)
|
||||
|
||||
# Process through preprocessor
|
||||
processed = preprocessor(batch)
|
||||
|
||||
# Check that data stays on same GPU
|
||||
assert processed[OBS_STATE].device == device
|
||||
assert processed[OBS_IMAGE].device == device
|
||||
assert processed[TransitionKey.ACTION.value].device == device
|
||||
|
||||
|
||||
@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="Requires at least 2 GPUs")
|
||||
def test_pi0_processor_multi_gpu():
|
||||
"""Test PI0 processor with multi-GPU setup."""
|
||||
config = create_default_config()
|
||||
config.device = "cuda:0"
|
||||
stats = create_default_stats()
|
||||
|
||||
# Mock the tokenizer processor to act as pass-through
|
||||
class MockTokenizerProcessorStep(ProcessorStep):
|
||||
def __init__(self, *args, **kwargs):
|
||||
pass
|
||||
|
||||
def __call__(self, transition):
|
||||
return transition
|
||||
|
||||
def state_dict(self):
|
||||
return {}
|
||||
|
||||
def load_state_dict(self, state):
|
||||
pass
|
||||
|
||||
def reset(self):
|
||||
pass
|
||||
|
||||
def get_config(self):
|
||||
return {"tokenizer_name": "google/paligemma-3b-pt-224"}
|
||||
|
||||
def transform_features(self, features):
|
||||
return features
|
||||
|
||||
with patch("lerobot.policies.pi0.processor_pi0.TokenizerProcessorStep", MockTokenizerProcessorStep):
|
||||
preprocessor, postprocessor = make_pi0_pre_post_processors(
|
||||
config,
|
||||
stats,
|
||||
)
|
||||
|
||||
# Simulate data on different GPU
|
||||
device = torch.device("cuda:1")
|
||||
observation = {
|
||||
OBS_STATE: torch.randn(1, 10).to(device),
|
||||
OBS_IMAGE: torch.randn(1, 3, 224, 224).to(device),
|
||||
}
|
||||
action = torch.randn(1, 6).to(device)
|
||||
transition = create_transition(observation, action, complementary_data={"task": ["test task"]})
|
||||
batch = transition_to_batch(transition)
|
||||
|
||||
# Process through preprocessor
|
||||
processed = preprocessor(batch)
|
||||
|
||||
# Check that data stays on cuda:1
|
||||
assert processed[OBS_STATE].device == device
|
||||
assert processed[OBS_IMAGE].device == device
|
||||
assert processed[TransitionKey.ACTION.value].device == device
|
||||
|
||||
|
||||
def test_pi0_processor_without_stats():
|
||||
"""Test PI0 processor creation without dataset statistics."""
|
||||
config = create_default_config()
|
||||
|
||||
# Mock the tokenizer processor
|
||||
with patch("lerobot.policies.pi0.processor_pi0.TokenizerProcessorStep", MockTokenizerProcessorStep):
|
||||
preprocessor, postprocessor = make_pi0_pre_post_processors(
|
||||
config,
|
||||
dataset_stats=None,
|
||||
)
|
||||
|
||||
# Should still create processors
|
||||
assert preprocessor is not None
|
||||
assert postprocessor is not None
|
||||
|
||||
|
||||
def test_pi0_newline_processor_state_dict():
|
||||
"""Test Pi0NewLineProcessor state dict methods."""
|
||||
processor = Pi0NewLineProcessor()
|
||||
|
||||
# Test state_dict (should be empty)
|
||||
state = processor.state_dict()
|
||||
assert state == {}
|
||||
|
||||
# Test load_state_dict (should do nothing)
|
||||
processor.load_state_dict({})
|
||||
|
||||
# Test reset (should do nothing)
|
||||
processor.reset()
|
||||
|
||||
# Test get_config
|
||||
config = processor.get_config()
|
||||
assert config == {}
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
|
||||
def test_pi0_processor_bfloat16_device_float32_normalizer():
|
||||
"""Test: DeviceProcessor(bfloat16) + NormalizerProcessor(float32) → output bfloat16 via automatic adaptation"""
|
||||
config = create_default_config()
|
||||
stats = create_default_stats()
|
||||
config.device = "cuda"
|
||||
|
||||
with patch("lerobot.policies.pi0.processor_pi0.TokenizerProcessorStep", MockTokenizerProcessorStep):
|
||||
preprocessor, _ = make_pi0_pre_post_processors(
|
||||
config,
|
||||
stats,
|
||||
)
|
||||
|
||||
# Modify the pipeline to use bfloat16 device processor with float32 normalizer
|
||||
modified_steps = []
|
||||
for step in preprocessor.steps:
|
||||
if isinstance(step, DeviceProcessorStep):
|
||||
# Device processor converts to bfloat16
|
||||
modified_steps.append(DeviceProcessorStep(device=config.device, float_dtype="bfloat16"))
|
||||
elif isinstance(step, NormalizerProcessorStep):
|
||||
# Normalizer stays configured as float32 (will auto-adapt to bfloat16)
|
||||
norm_step = step # Now type checker knows this is NormalizerProcessorStep
|
||||
modified_steps.append(
|
||||
NormalizerProcessorStep(
|
||||
features=norm_step.features,
|
||||
norm_map=norm_step.norm_map,
|
||||
stats=norm_step.stats,
|
||||
device=config.device,
|
||||
dtype=torch.float32, # Deliberately configured as float32
|
||||
)
|
||||
)
|
||||
else:
|
||||
modified_steps.append(step)
|
||||
preprocessor.steps = modified_steps
|
||||
|
||||
# Verify initial normalizer configuration (PI0 has NormalizerProcessorStep at index 5)
|
||||
normalizer_step = preprocessor.steps[5] # NormalizerProcessorStep
|
||||
assert normalizer_step.dtype == torch.float32
|
||||
|
||||
# Create test data with both state and visual observations
|
||||
observation = {
|
||||
OBS_STATE: torch.randn(10, dtype=torch.float32), # PI0 expects size 10
|
||||
OBS_IMAGE: torch.randn(3, 224, 224, dtype=torch.float32),
|
||||
}
|
||||
action = torch.randn(6, dtype=torch.float32) # PI0 expects size 6
|
||||
transition = create_transition(
|
||||
observation, action, complementary_data={"task": "test bfloat16 adaptation"}
|
||||
)
|
||||
batch = transition_to_batch(transition)
|
||||
|
||||
# Process through full pipeline
|
||||
processed = preprocessor(batch)
|
||||
|
||||
# Verify: DeviceProcessor → bfloat16, NormalizerProcessor adapts → final output is bfloat16
|
||||
assert processed[OBS_STATE].dtype == torch.bfloat16
|
||||
assert processed[OBS_IMAGE].dtype == torch.bfloat16 # IDENTITY normalization still gets dtype conversion
|
||||
assert processed[TransitionKey.ACTION.value].dtype == torch.bfloat16
|
||||
|
||||
# Verify normalizer automatically adapted its internal state
|
||||
assert normalizer_step.dtype == torch.bfloat16
|
||||
# Check state stats (has normalization)
|
||||
for stat_tensor in normalizer_step._tensor_stats[OBS_STATE].values():
|
||||
assert stat_tensor.dtype == torch.bfloat16
|
||||
# OBS_IMAGE uses IDENTITY normalization, so no stats to check
|
||||
Reference in New Issue
Block a user