careamics 0.0.15__py3-none-any.whl → 0.0.17__py3-none-any.whl

careamics/lightning/microsplit_data_module.py

@@ -0,0 +1,636 @@
+"""MicroSplit data module for training and validation."""
+
+from collections.abc import Callable
+from pathlib import Path
+from typing import Union
+
+import numpy as np
+import pytorch_lightning as L
+import tifffile
+from numpy.typing import NDArray
+from torch.utils.data import DataLoader
+
+from careamics.dataset.dataset_utils.dataset_utils import reshape_array
+from careamics.lvae_training.dataset import (
+    DataSplitType,
+    DataType,
+    LCMultiChDloader,
+    MicroSplitDataConfig,
+)
+from careamics.lvae_training.dataset.types import TilingMode
+
+
+# TODO refactor
+def load_one_file(fpath):
+    """Load a single 2D image file.
+
+    Parameters
+    ----------
+    fpath : str or Path
+        Path to the image file.
+
+    Returns
+    -------
+    numpy.ndarray
+        Reshaped image data.
+    """
+    data = tifffile.imread(fpath)
+    if len(data.shape) == 2:
+        axes = "YX"
+    elif len(data.shape) == 3:
+        axes = "SYX"
+    elif len(data.shape) == 4:
+        axes = "STYX"
+    else:
+        raise ValueError(f"Invalid data shape: {data.shape}")
+    data = reshape_array(data, axes)
+    data = data.reshape(-1, data.shape[-2], data.shape[-1])
+    return data
+
+
+# TODO refactor
+def load_data(datadir):
+    """Load data from a directory containing channel subdirectories with image files.
+
+    Parameters
+    ----------
+    datadir : str or Path
+        Path to the data directory containing channel subdirectories.
+
+    Returns
+    -------
+    numpy.ndarray
+        Stacked array of all channels' data.
+    """
+    data_path = Path(datadir)
+
+    channel_dirs = sorted(p for p in data_path.iterdir() if p.is_dir())
+    channels_data = []
+
+    for channel_dir in channel_dirs:
+        image_files = sorted(f for f in channel_dir.iterdir() if f.is_file())
+        channel_images = [load_one_file(image_path) for image_path in image_files]
+
+        channel_stack = np.concatenate(
+            channel_images, axis=0
+        )  # FIXME: this line works iff images have
+        # a singleton channel dimension. Specify in the notebook or change with
+        # `torch.stack`??
+        channels_data.append(channel_stack)
+
+    final_data = np.stack(channels_data, axis=-1)
+    return final_data
+
+
+# TODO refactor
+def get_datasplit_tuples(val_fraction, test_fraction, data_length):
+    """Get train/val/test indices for data splitting.
+
+    Parameters
+    ----------
+    val_fraction : float or None
+        Fraction of data to use for validation.
+    test_fraction : float or None
+        Fraction of data to use for testing.
+    data_length : int
+        Total length of the dataset.
+
+    Returns
+    -------
+    tuple[numpy.ndarray, numpy.ndarray, numpy.ndarray]
+        Training, validation, and test indices.
+    """
+    indices = np.arange(data_length)
+    np.random.shuffle(indices)
+
+    if val_fraction is None:
+        val_fraction = 0.0
+    if test_fraction is None:
+        test_fraction = 0.0
+
+    val_size = int(data_length * val_fraction)
+    test_size = int(data_length * test_fraction)
+    train_size = data_length - val_size - test_size
+
+    train_idx = indices[:train_size]
+    val_idx = indices[train_size : train_size + val_size]
+    test_idx = indices[train_size + val_size :]
+
+    return train_idx, val_idx, test_idx
+
+
+# TODO refactor
+def get_train_val_data(
+    data_config,
+    datadir,
+    datasplit_type: DataSplitType,
+    val_fraction=None,
+    test_fraction=None,
+    allow_generation=None,
+    **kwargs,
+):
+    """Load and split data according to configuration.
+
+    Parameters
+    ----------
+    data_config : MicroSplitDataConfig
+        Data configuration object.
+    datadir : str or Path
+        Path to the data directory.
+    datasplit_type : DataSplitType
+        Type of data split to return.
+    val_fraction : float, optional
+        Fraction of data to use for validation.
+    test_fraction : float, optional
+        Fraction of data to use for testing.
+    allow_generation : bool, optional
+        Whether to allow data generation.
+    **kwargs
+        Additional keyword arguments.
+
+    Returns
+    -------
+    numpy.ndarray
+        Split data array.
+    """
+    data = load_data(datadir)
+    train_idx, val_idx, test_idx = get_datasplit_tuples(
+        val_fraction, test_fraction, len(data)
+    )
+
+    if datasplit_type == DataSplitType.All:
+        data = data.astype(np.float64)
+    elif datasplit_type == DataSplitType.Train:
+        data = data[train_idx].astype(np.float64)
+    elif datasplit_type == DataSplitType.Val:
+        data = data[val_idx].astype(np.float64)
+    elif datasplit_type == DataSplitType.Test:
+        # TODO this is only used for prediction, and only because old dataset uses it
+        data = data[test_idx].astype(np.float64)
+    else:
+        raise Exception("invalid datasplit")
+
+    return data
+
+
+class MicroSplitDataModule(L.LightningDataModule):
+    """Lightning DataModule for MicroSplit-style datasets.
+
+    Matches the interface of TrainDataModule, but internally uses original MicroSplit
+    dataset logic.
+
+    Parameters
+    ----------
+    data_config : MicroSplitDataConfig
+        Configuration for the MicroSplit dataset.
+    train_data : str
+        Path to training data directory.
+    val_data : str, optional
+        Path to validation data directory.
+    train_data_target : str, optional
+        Path to training target data.
+    val_data_target : str, optional
+        Path to validation target data.
+    read_source_func : Callable, optional
+        Function to read source data.
+    extension_filter : str, optional
+        File extension filter.
+    val_percentage : float, optional
+        Percentage of data to use for validation, by default 0.1.
+    val_minimum_split : int, optional
+        Minimum number of samples for validation split, by default 5.
+    use_in_memory : bool, optional
+        Whether to use in-memory dataset, by default True.
+    """
+
+    def __init__(
+        self,
+        # Should be compatible with microSplit DatasetConfig
+        data_config: MicroSplitDataConfig,
+        train_data: str,
+        val_data: str | None = None,
+        train_data_target: str | None = None,
+        val_data_target: str | None = None,
+        read_source_func: Callable | None = None,
+        extension_filter: str = "",
+        val_percentage: float = 0.1,
+        val_minimum_split: int = 5,
+        use_in_memory: bool = True,
+    ):
+        """Initialize MicroSplitDataModule.
+
+        Parameters
+        ----------
+        data_config : MicroSplitDataConfig
+            Configuration for the MicroSplit dataset.
+        train_data : str
+            Path to training data directory.
+        val_data : str, optional
+            Path to validation data directory.
+        train_data_target : str, optional
+            Path to training target data.
+        val_data_target : str, optional
+            Path to validation target data.
+        read_source_func : Callable, optional
+            Function to read source data.
+        extension_filter : str, optional
+            File extension filter.
+        val_percentage : float, optional
+            Percentage of data to use for validation, by default 0.1.
+        val_minimum_split : int, optional
+            Minimum number of samples for validation split, by default 5.
+        use_in_memory : bool, optional
+            Whether to use in-memory dataset, by default True.
+        """
+        super().__init__()
+        # Dataset selection logic (adapted from create_train_val_datasets)
+        self.train_config = data_config  # SHould configs be separated?
+        self.val_config = data_config
+        self.test_config = data_config
+
+        datapath = train_data
+        load_data_func = read_source_func
+
+        dataset_class = LCMultiChDloader  # TODO hardcoded for now
+
+        # Create datasets
+        self.train_dataset = dataset_class(
+            self.train_config,
+            datapath,
+            load_data_fn=load_data_func,
+            val_fraction=val_percentage,
+            test_fraction=0.1,
+        )
+        max_val = self.train_dataset.get_max_val()
+        self.val_config.max_val = max_val
+        if self.train_config.datasplit_type == DataSplitType.All:
+            self.val_config.datasplit_type = DataSplitType.All
+            self.test_config.datasplit_type = DataSplitType.All
+        self.val_dataset = dataset_class(
+            self.val_config,
+            datapath,
+            load_data_fn=load_data_func,
+            val_fraction=val_percentage,
+            test_fraction=0.1,
+        )
+        self.test_config.max_val = max_val
+        self.test_dataset = dataset_class(
+            self.test_config,
+            datapath,
+            load_data_fn=load_data_func,
+            val_fraction=val_percentage,
+            test_fraction=0.1,
+        )
+        mean_val, std_val = self.train_dataset.compute_mean_std()
+        self.train_dataset.set_mean_std(mean_val, std_val)
+        self.val_dataset.set_mean_std(mean_val, std_val)
+        self.test_dataset.set_mean_std(mean_val, std_val)
+        data_stats = self.train_dataset.get_mean_std()
+
+        # Store data statistics
+        self.data_stats = (
+            data_stats[0],
+            data_stats[1],
+        )  # TODO repeats old logic, revisit
+
+    def train_dataloader(self):
+        """Create a dataloader for training.
+
+        Returns
+        -------
+        DataLoader
+            Training dataloader.
+        """
+        return DataLoader(
+            self.train_dataset,
+            # TODO should be inside dataloader params?
+            batch_size=self.train_config.batch_size,
+            **self.train_config.train_dataloader_params,
+        )
+
+    def val_dataloader(self):
+        """Create a dataloader for validation.
+
+        Returns
+        -------
+        DataLoader
+            Validation dataloader.
+        """
+        return DataLoader(
+            self.val_dataset,
+            batch_size=self.train_config.batch_size,
+            **self.val_config.val_dataloader_params,  # TODO duplicated
+        )
+
+    def get_data_stats(self):
+        """Get data statistics.
+
+        Returns
+        -------
+        tuple[dict, dict]
+            A tuple containing two dictionaries:
+            - data_mean: mean values for input and target
+            - data_std: standard deviation values for input and target
+        """
+        return self.data_stats, self.val_config.max_val  # TODO should be in the config?
+
+
+def create_microsplit_train_datamodule(
+    train_data: str,
+    patch_size: tuple,
+    data_type: DataType,
+    axes: str,  # TODO should be there after refactoring
+    batch_size: int,
+    val_data: str | None = None,
+    num_channels: int = 2,
+    depth3D: int = 1,
+    grid_size: tuple | None = None,
+    multiscale_count: int | None = None,
+    tiling_mode: TilingMode = TilingMode.ShiftBoundary,
+    read_source_func: Callable | None = None,  # TODO should be there after refactoring
+    extension_filter: str = "",
+    val_percentage: float = 0.1,
+    val_minimum_split: int = 5,
+    use_in_memory: bool = True,
+    transforms: list | None = None,  # TODO should it be here?
+    train_dataloader_params: dict | None = None,
+    val_dataloader_params: dict | None = None,
+    **dataset_kwargs,
+) -> MicroSplitDataModule:
+    """
+    Create a MicroSplitDataModule for microSplit-style datasets.
+
+    This includes config creation.
+
+    Parameters
+    ----------
+    train_data : str
+        Path to training data.
+    patch_size : tuple
+        Size of one patch of data.
+    data_type : DataType
+        Type of the dataset (must be a DataType enum value).
+    axes : str
+        Axes of the data (e.g., 'SYX').
+    batch_size : int
+        Batch size for dataloaders.
+    val_data : str, optional
+        Path to validation data.
+    num_channels : int, default=2
+        Number of channels in the input.
+    depth3D : int, default=1
+        Number of slices in 3D.
+    grid_size : tuple, optional
+        Grid size for patch extraction.
+    multiscale_count : int, optional
+        Number of LC scales.
+    tiling_mode : TilingMode, default=ShiftBoundary
+        Tiling mode for patch extraction.
+    read_source_func : Callable, optional
+        Function to read the source data.
+    extension_filter : str, optional
+        File extension filter.
+    val_percentage : float, default=0.1
+        Percentage of training data to use for validation.
+    val_minimum_split : int, default=5
+        Minimum number of patches/files for validation split.
+    use_in_memory : bool, default=True
+        Use in-memory dataset if possible.
+    transforms : list, optional
+        List of transforms to apply.
+    train_dataloader_params : dict, optional
+        Parameters for training dataloader.
+    val_dataloader_params : dict, optional
+        Parameters for validation dataloader.
+    **dataset_kwargs :
+        Additional arguments passed to DatasetConfig.
+
+    Returns
+    -------
+    MicroSplitDataModule
+        Configured MicroSplitDataModule instance.
+    """
+    # Create dataset configs with only valid parameters
+    dataset_config_params = {
+        "data_type": data_type,
+        "image_size": patch_size,
+        "num_channels": num_channels,
+        "depth3D": depth3D,
+        "grid_size": grid_size,
+        "multiscale_lowres_count": multiscale_count,
+        "tiling_mode": tiling_mode,
+        "batch_size": batch_size,
+        "train_dataloader_params": train_dataloader_params,
+        "val_dataloader_params": val_dataloader_params,
+        **dataset_kwargs,
+    }
+
+    train_config = MicroSplitDataConfig(
+        **dataset_config_params,
+        datasplit_type=DataSplitType.Train,
+    )
+    # val_config = MicroSplitDataConfig(
+    #     **dataset_config_params,
+    #     datasplit_type=DataSplitType.Val,
+    # )
+    # TODO, data config is duplicated here and in configuration
+
+    return MicroSplitDataModule(
+        data_config=train_config,
+        train_data=train_data,
+        val_data=val_data or train_data,
+        train_data_target=None,
+        val_data_target=None,
+        read_source_func=get_train_val_data,  # Use our wrapped function
+        extension_filter=extension_filter,
+        val_percentage=val_percentage,
+        val_minimum_split=val_minimum_split,
+        use_in_memory=use_in_memory,
+    )
+
+
+class MicroSplitPredictDataModule(L.LightningDataModule):
+    """Lightning DataModule for MicroSplit-style prediction datasets.
+
+    Matches the interface of PredictDataModule, but internally uses MicroSplit
+    dataset logic for prediction.
+
+    Parameters
+    ----------
+    pred_config : MicroSplitDataConfig
+        Configuration for MicroSplit prediction.
+    pred_data : str or Path or numpy.ndarray
+        Prediction data, can be a path to a folder, a file or a numpy array.
+    read_source_func : Callable, optional
+        Function to read custom types.
+    extension_filter : str, optional
+        Filter to filter file extensions for custom types.
+    dataloader_params : dict, optional
+        Dataloader parameters.
+    """
+
+    def __init__(
+        self,
+        pred_config: MicroSplitDataConfig,
+        pred_data: Union[str, Path, NDArray],
+        read_source_func: Callable | None = None,
+        extension_filter: str = "",
+        dataloader_params: dict | None = None,
+    ) -> None:
+        """
+        Constructor for MicroSplit prediction data module.
+
+        Parameters
+        ----------
+        pred_config : MicroSplitDataConfig
+            Configuration for MicroSplit prediction.
+        pred_data : str or Path or numpy.ndarray
+            Prediction data, can be a path to a folder, a file or a numpy array.
+        read_source_func : Callable, optional
+            Function to read custom types, by default None.
+        extension_filter : str, optional
+            Filter to filter file extensions for custom types, by default "".
+        dataloader_params : dict, optional
+            Dataloader parameters, by default {}.
+        """
+        super().__init__()
+
+        if dataloader_params is None:
+            dataloader_params = {}
+        self.pred_config = pred_config
+        self.pred_data = pred_data
+        self.read_source_func = read_source_func or get_train_val_data
+        self.extension_filter = extension_filter
+        self.dataloader_params = dataloader_params
+
+    def prepare_data(self) -> None:
+        """Hook used to prepare the data before calling `setup`."""
+        # # TODO currently data preparation is handled in dataset creation, revisit!
+        pass
+
+    def setup(self, stage: str | None = None) -> None:
+        """
+        Hook called at the beginning of predict.
+
+        Parameters
+        ----------
+        stage : Optional[str], optional
+            Stage, by default None.
+        """
+        # Create prediction dataset using LCMultiChDloader
+        self.predict_dataset = LCMultiChDloader(
+            self.pred_config,
+            self.pred_data,
+            load_data_fn=self.read_source_func,
+            val_fraction=0.0,  # No validation split for prediction
+            test_fraction=1.0,  # No test split for prediction
+        )
+        self.predict_dataset.set_mean_std(*self.pred_config.data_stats)
+
+    def predict_dataloader(self) -> DataLoader:
+        """
+        Create a dataloader for prediction.
+
+        Returns
+        -------
+        DataLoader
+            Prediction dataloader.
+        """
+        return DataLoader(
+            self.predict_dataset,
+            batch_size=self.pred_config.batch_size,
+            **self.dataloader_params,
+        )
+
+
+def create_microsplit_predict_datamodule(
+    pred_data: Union[str, Path, NDArray],
+    tile_size: tuple,
+    data_type: DataType,
+    axes: str,
+    batch_size: int = 1,
+    num_channels: int = 2,
+    depth3D: int = 1,
+    grid_size: int | None = None,
+    multiscale_count: int | None = None,
+    data_stats: tuple | None = None,
+    tiling_mode: TilingMode = TilingMode.ShiftBoundary,
+    read_source_func: Callable | None = None,
+    extension_filter: str = "",
+    dataloader_params: dict | None = None,
+    **dataset_kwargs,
+) -> MicroSplitPredictDataModule:
+    """
+    Create a MicroSplitPredictDataModule for microSplit-style prediction datasets.
+
+    Parameters
+    ----------
+    pred_data : str or Path or numpy.ndarray
+        Prediction data, can be a path to a folder, a file or a numpy array.
+    tile_size : tuple
+        Size of one tile of data.
+    data_type : DataType
+        Type of the dataset (must be a DataType enum value).
+    axes : str
+        Axes of the data (e.g., 'SYX').
+    batch_size : int, default=1
+        Batch size for prediction dataloader.
+    num_channels : int, default=2
+        Number of channels in the input.
+    depth3D : int, default=1
+        Number of slices in 3D.
+    grid_size : tuple, optional
+        Grid size for patch extraction.
+    multiscale_count : int, optional
+        Number of LC scales.
+    data_stats : tuple, optional
+        Data statistics, by default None.
+    tiling_mode : TilingMode, default=ShiftBoundary
+        Tiling mode for patch extraction.
+    read_source_func : Callable, optional
+        Function to read the source data.
+    extension_filter : str, optional
+        File extension filter.
+    dataloader_params : dict, optional
+        Parameters for prediction dataloader.
+    **dataset_kwargs :
+        Additional arguments passed to MicroSplitDataConfig.
+
+    Returns
+    -------
+    MicroSplitPredictDataModule
+        Configured MicroSplitPredictDataModule instance.
+    """
+    if dataloader_params is None:
+        dataloader_params = {}
+
+    # Create prediction config with only valid parameters
+    prediction_config_params = {
+        "data_type": data_type,
+        "image_size": tile_size,
+        "num_channels": num_channels,
+        "depth3D": depth3D,
+        "grid_size": grid_size,
+        "multiscale_lowres_count": multiscale_count,
+        "data_stats": data_stats,
+        "tiling_mode": tiling_mode,
+        "batch_size": batch_size,
+        "datasplit_type": DataSplitType.Test,  # For prediction, use all data
+        **dataset_kwargs,
+    }
+
+    pred_config = MicroSplitDataConfig(**prediction_config_params)
+
+    # Remove batch_size from dataloader_params if present
+    if "batch_size" in dataloader_params:
+        del dataloader_params["batch_size"]
+
+    return MicroSplitPredictDataModule(
+        pred_config=pred_config,
+        pred_data=pred_data,
+        read_source_func=(
+            read_source_func if read_source_func is not None else get_train_val_data
+        ),
+        extension_filter=extension_filter,
+        dataloader_params=dataloader_params,
+    )

careamics/lightning/predict_data_module.py

@@ -217,11 +217,18 @@ class PredictDataModule(L.LightningDataModule):
         DataLoader
             Prediction dataloader.
         """
+        # For tiled predictions, we need to ensure tiles are processed in order
+        # to avoid stitching artifacts. Multi-worker processing can return batches
+        # out of order, so we disable it for tiled predictions.
+        dataloader_params = self.dataloader_params.copy()
+        if self.tiled:
+            dataloader_params["num_workers"] = 0
+
         return DataLoader(
             self.predict_dataset,
             batch_size=self.batch_size,
             collate_fn=collate_tiles if self.tiled else None,
-            **self.dataloader_params,
+            **dataloader_params,
         )