Add data module for LEVIR-CD+ dataset

tests/datamodules/test_levircd.py

@@ -0,0 +1,61 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+import os
+
+import pytest
+from _pytest.fixtures import SubRequest
+from lightning.pytorch import Trainer
+
+from torchgeo.datamodules import LEVIRCDPlusDataModule
+from torchgeo.datasets import LEVIRCDPlus
+
+
+class TestLEVIRCDPlusDataModule:
+    def datamodule(self, request: SubRequest) -> OSCDDataModule:
+        bands = request.param
+        root = os.path.join("tests", "data", "LEVIR-CD+")
+        dm = LEVIRCDPlusDataModule(root=root, download=True, num_workers=0)
+        dm.prepare_data()
+        dm.trainer = Trainer(accelerator="cpu", max_epochs=1)
+        return dm
+
+    def test_train_dataloader(self, datamodule: LEVIRCDPlusDataModule) -> None:
+        datamodule.setup("fit")
+        if datamodule.trainer:
+            datamodule.trainer.training = True
+        batch = next(iter(datamodule.train_dataloader()))
+        batch = datamodule.on_after_batch_transfer(batch, 0)
+        assert batch["image1"].shape[-2:] == batch["mask"].shape[-2:] == (2, 2)
+        assert batch["image1"].shape[0] == batch["mask"].shape[0] == 1
+        assert batch["image2"].shape[-2:] == batch["mask"].shape[-2:] == (2, 2)
+        assert batch["image2"].shape[0] == batch["mask"].shape[0] == 1
+        assert batch["image1"].shape[1] == 3
+        assert batch["image2"].shape[1] == 3
+
+    def test_val_dataloader(self, datamodule: LEVIRCDPlusDataModule) -> None:
+        datamodule.setup("validate")
+        if datamodule.trainer:
+            datamodule.trainer.validating = True
+        batch = next(iter(datamodule.val_dataloader()))
+        batch = datamodule.on_after_batch_transfer(batch, 0)
+        if datamodule.val_split_pct > 0.0:
+            assert batch["image1"].shape[-2:] == batch["mask"].shape[-2:] == (2, 2)
+            assert batch["image1"].shape[0] == batch["mask"].shape[0] == 1
+            assert batch["image2"].shape[-2:] == batch["mask"].shape[-2:] == (2, 2)
+            assert batch["image2"].shape[0] == batch["mask"].shape[0] == 1
+            assert batch["image1"].shape[1] == 3
+            assert batch["image2"].shape[1] == 3
+
+    def test_test_dataloader(self, datamodule: LEVIRCDPlusDataModule) -> None:
+        datamodule.setup("test")
+        if datamodule.trainer:
+            datamodule.trainer.testing = True
+        batch = next(iter(datamodule.test_dataloader()))
+        batch = datamodule.on_after_batch_transfer(batch, 0)
+        assert batch["image1"].shape[-2:] == batch["mask"].shape[-2:] == (2, 2)
+        assert batch["image1"].shape[0] == batch["mask"].shape[0] == 1
+        assert batch["image2"].shape[-2:] == batch["mask"].shape[-2:] == (2, 2)
+        assert batch["image2"].shape[0] == batch["mask"].shape[0] == 1
+        assert batch["image1"].shape[1] == 3
+        assert batch["image2"].shape[1] == 3

torchgeo/datamodules/__init__.py

@@ -18,6 +18,7 @@
 from .l7irish import L7IrishDataModule
 from .l8biome import L8BiomeDataModule
 from .landcoverai import LandCoverAIDataModule
+from .levircd import LEVIRCDPlusDataModule
 from .loveda import LoveDADataModule
 from .naip import NAIPChesapeakeDataModule
 from .nasa_marine_debris import NASAMarineDebrisDataModule

torchgeo/datamodules/levircd.py

@@ -0,0 +1,70 @@
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+
+"""LEVIR-CD+ datamodule."""
+
+from typing import Union
+
+import kornia.augmentation as K
+import torch
+
+from torchgeo.datamodules.utils import dataset_split
+from torchgeo.samplers.utils import _to_tuple
+
+from ..datasets import LEVIRCDPlus
+from ..transforms import AugmentationSequential
+from ..transforms.transforms import _RandomNCrop
+from .geo import NonGeoDataModule
+
+
+class LEVIRCDPlusDataModule(NonGeoDataModule):
+    """LightningDataModule implementation for the LEVIR-CD+ dataset.
+
+    Uses the train/test splits from the dataset and further splits
+    the train split into train/val splits.
+
+    """
+
+    def __init__(
+        self,
+        batch_size: int = 8,
+        patch_size: Union[tuple[int, int], int] = 256,
+        val_split_pct: float = 0.2,
+        num_workers: int = 0,
+        **kwargs,
+    ) -> None:
+        """Initialize a new LEVIRCDPlusDataModule instance.
+
+        Args:
+            batch_size: Size of each mini-batch.
+            patch_size: Size of each patch, either ``size`` or ``(height, width)``.
+                Should be a multiple of 32 for most segmentation architectures.
+            val_split_pct: Percentage of the dataset to use as a validation set.
+            num_workers: Number of workers for parallel data loading.
+            **kwargs: Additional keyword arguments passed to
+                :class:`~torchgeo.datasets.LEVIRCDPlus`.
+        """
+        super().__init__(LEVIRCDPlusDataModule, 1, num_workers, **kwargs)
+
+        self.patch_size = _to_tuple(patch_size)
+        self.val_split_pct = val_split_pct
+
+        self.aug = AugmentationSequential(
+            K.Normalize(mean=self.mean, std=self.std),
+            _RandomNCrop(self.patch_size, batch_size),
+            data_keys=["image1", "image2", "mask"],
+        )
+
+    def setup(self, stage: str) -> None:
+        """Set up datasets.
+
+        Args:
+            stage: Either 'fit', 'validate', 'test', or 'predict'.
+        """
+        if stage in ["fit", "validate"]:
+            self.dataset = LEVIRCDPlus(split="train", **self.kwargs)
+            self.train_dataset, self.val_dataset, _ = dataset_split(
+                self.dataset, val_pct=self.val_split_pct, test_pct=0
+            )
+        if stage in ["test"]:
+            self.test_dataset = LEVIRCDPlus(split="test", **self.kwargs)