diff --git a/nerfstudio/pipelines/base_pipeline.py b/nerfstudio/pipelines/base_pipeline.py index 9d062e02f..7b35d98cc 100644 --- a/nerfstudio/pipelines/base_pipeline.py +++ b/nerfstudio/pipelines/base_pipeline.py @@ -21,7 +21,7 @@ import typing from abc import abstractmethod from dataclasses import dataclass, field from time import time -from typing import Any, Dict, List, Optional, Type, Union, cast +from typing import Any, Dict, List, Mapping, Optional, Type, Union, cast import torch import torch.distributed as dist @@ -105,6 +105,16 @@ class Pipeline(nn.Module): """Returns the device that the model is on.""" return self.model.device + def load_state_dict(self, state_dict: Mapping[str, Any], strict: bool = True): + if "model" in state_dict: + tmp_state_dict = dict(state_dict) + model_state = tmp_state_dict.pop("model") + self.model.load_state_dict(model_state, strict=strict) + + # Model will be missing from the dict; therefore strict=False here. + strict = False + super().load_state_dict(state_dict, strict=strict) + @profiler.time_function def get_train_loss_dict(self, step: int): """This function gets your training loss dict. This will be responsible for @@ -362,7 +372,10 @@ class VanillaPipeline(Pipeline): """ state = {key.replace("module.", ""): value for key, value in loaded_state.items()} self._model.update_to_step(step) + model_state = state.pop("model", {}) + self.model.load_state_dict(model_state, strict=True) self.load_state_dict(state, strict=False) + state["model"] = model_state def get_training_callbacks( self, training_callback_attributes: TrainingCallbackAttributes diff --git a/tests/pipelines/test_vanilla_pipeline.py b/tests/pipelines/test_vanilla_pipeline.py new file mode 100644 index 000000000..64168c300 --- /dev/null +++ b/tests/pipelines/test_vanilla_pipeline.py @@ -0,0 +1,71 @@ +""" +Test pipeline +""" +from pathlib import Path + +# pylint: disable=too-few-public-methods +# pylint: disable=no-self-use +# pylint: disable=missing-class-docstring +# pylint: disable=unused-argument +# pylint: disable=abstract-method +import torch + +from nerfstudio.cameras.cameras import Cameras +from nerfstudio.data.datasets.base_dataset import DataparserOutputs, InputDataset +from nerfstudio.pipelines.base_pipeline import ( + Model, + ModelConfig, + VanillaDataManagerConfig, + VanillaPipeline, + VanillaPipelineConfig, +) + + +class MockedDataManager: + """Mocked data manager""" + + def __init__(self, *args, **kwargs): + num_images = 0 + self.train_dataset = InputDataset( + DataparserOutputs( + image_filenames=[Path("filename.png")] * num_images, + cameras=Cameras( + camera_to_worlds=torch.ones([num_images, 3, 4], dtype=torch.float32), + fx=1.0, + fy=1.0, + cx=1.0, + cy=1.0, + width=2, + height=2, + ), + ) + ) + + def to(self, *args, **kwargs): + "Mocked to" + return self + + +def test_load_state_dict(): + """Test pipeline load_state_dict calls model's load_state_dict""" + was_called = False + + class MockedModel(Model): # + """Mocked model""" + + def load_state_dict(self, *args, strict=True): + """Mocked load_state_dict""" + nonlocal was_called + was_called = True + assert strict + + config = VanillaPipelineConfig( + datamanager=VanillaDataManagerConfig( + _target=MockedDataManager, + ), + model=ModelConfig(_target=MockedModel), + ) + pipeline = VanillaPipeline(config, "cpu") + state_dict = pipeline.state_dict() + pipeline.load_pipeline(state_dict, 0) + assert was_called