rslearn 0.0.12__py3-none-any.whl → 0.0.14__py3-none-any.whl

rslearn/train/all_patches_dataset.py

@@ -0,0 +1,458 @@
+"""Wrapper around ModelDataset to load all patches (crops) in a window."""
+
+import itertools
+from collections.abc import Iterable, Iterator
+from typing import Any
+
+import shapely
+import torch
+
+from rslearn.dataset import Window
+from rslearn.train.dataset import ModelDataset
+from rslearn.utils.geometry import PixelBounds, STGeometry
+
+
+def get_window_patch_options(
+    patch_size: tuple[int, int],
+    overlap_size: tuple[int, int],
+    bounds: PixelBounds,
+) -> list[PixelBounds]:
+    """Get the bounds of each input patch within the window bounds.
+
+    This is used when running inference on all patches (crops) of a large window, to
+    compute the position of each patch.
+
+    Args:
+        patch_size: the size of the patches to extract.
+        overlap_size: the size of the overlap between patches.
+        bounds: the window bounds to divide up into smaller patches.
+
+    Returns:
+        a list of patch bounds within the overall bounds. The rightmost and
+            bottommost patches may extend beyond the provided bounds.
+    """
+    # We stride the patches by patch_size - overlap_size until the last patch.
+    # We handle the last patch with a special case to ensure it does not exceed the
+    # window bounds. Instead, it may overlap the previous patch.
+    cols = list(
+        range(
+            bounds[0],
+            bounds[2] - patch_size[0],
+            patch_size[0] - overlap_size[0],
+        )
+    ) + [bounds[2] - patch_size[0]]
+    rows = list(
+        range(
+            bounds[1],
+            bounds[3] - patch_size[1],
+            patch_size[1] - overlap_size[1],
+        )
+    ) + [bounds[3] - patch_size[1]]
+
+    patch_bounds: list[PixelBounds] = []
+    for col in cols:
+        for row in rows:
+            patch_bounds.append((col, row, col + patch_size[0], row + patch_size[1]))
+    return patch_bounds
+
+
+def pad_slice_protect(
+    raw_inputs: dict[str, Any],
+    passthrough_inputs: dict[str, Any],
+    patch_size: tuple[int, int],
+) -> tuple[dict[str, Any], dict[str, Any]]:
+    """Pad tensors in-place by patch size to protect slicing near right/bottom edges.
+
+    Args:
+        raw_inputs: the raw inputs to pad.
+        passthrough_inputs: the passthrough inputs to pad.
+        patch_size: the size of the patches to extract.
+
+    Returns:
+        a tuple of (raw_inputs, passthrough_inputs).
+    """
+    for d in [raw_inputs, passthrough_inputs]:
+        for input_name, value in list(d.items()):
+            if not isinstance(value, torch.Tensor):
+                continue
+            d[input_name] = torch.nn.functional.pad(
+                value, pad=(0, patch_size[0], 0, patch_size[1])
+            )
+    return raw_inputs, passthrough_inputs
+
+
+class IterableAllPatchesDataset(torch.utils.data.IterableDataset):
+    """This wraps a ModelDataset to iterate over all patches in that dataset.
+
+    This should be used when SplitConfig.load_all_patches is enabled. The ModelDataset
+    is configured with no patch size (load entire windows), and the dataset is wrapped
+    in an AllPatchesDataset.
+
+    Similar to DistributedSampler, we add extra samples at each rank to ensure
+    consistent number of batches across all ranks.
+    """
+
+    def __init__(
+        self,
+        dataset: ModelDataset,
+        patch_size: tuple[int, int],
+        overlap_ratio: float = 0.0,
+        rank: int = 0,
+        world_size: int = 1,
+    ):
+        """Create a new IterableAllPatchesDataset.
+
+        Args:
+            dataset: the ModelDataset to wrap.
+            patch_size: the size of the patches to extract.
+            overlap_ratio: whether to include overlap between the patches. Note that
+                the right/bottom-most patches may still overlap since we ensure that
+                all patches are contained in the window bounds.
+            rank: the global rank of this train worker process.
+            world_size: the total number of train worker processes.
+        """
+        super().__init__()
+        self.dataset = dataset
+        self.patch_size = patch_size
+        self.overlap_size = (
+            round(self.patch_size[0] * overlap_ratio),
+            round(self.patch_size[1] * overlap_ratio),
+        )
+        self.rank = rank
+        self.world_size = world_size
+        self.windows = self.dataset.get_dataset_examples()
+
+    def set_name(self, name: str) -> None:
+        """Sets dataset name.
+
+        Args:
+            name: dataset name
+        """
+        self.dataset.set_name(name)
+
+    def get_window_num_patches(self, bounds: PixelBounds) -> int:
+        """Get the number of patches for these bounds.
+
+        This corresponds to the length of the list returned by get_patch_options.
+        """
+        num_cols = (
+            len(
+                range(
+                    bounds[0],
+                    bounds[2] - self.patch_size[0],
+                    self.patch_size[0] - self.overlap_size[0],
+                )
+            )
+            + 1
+        )
+        num_rows = (
+            len(
+                range(
+                    bounds[1],
+                    bounds[3] - self.patch_size[1],
+                    self.patch_size[1] - self.overlap_size[1],
+                )
+            )
+            + 1
+        )
+        return num_cols * num_rows
+
+    def _get_worker_iteration_data(self) -> tuple[Iterable[int], int]:
+        """Get the windows we should iterate over.
+
+        This is split both by training worker (self.rank) and data loader worker (via
+        get_worker_info).
+
+        We also compute the total number of samples that each data loader worker should
+        yield. This is important for DDP to ensure that all ranks see the same number
+        of batches.
+
+        Returns:
+            a tuple (window_ids, num_samples_per_worker).
+        """
+        # Figure out the total number of data loader workers and our worker ID.
+        worker_info = torch.utils.data.get_worker_info()
+        if worker_info is None:
+            worker_id = 0
+            num_workers = 1
+        else:
+            worker_id = worker_info.id
+            num_workers = worker_info.num_workers
+        global_worker_id = self.rank * num_workers + worker_id
+        global_num_workers = self.world_size * num_workers
+
+        # Split up the windows evenly among the workers.
+        # We compute this for all workers since we will need to see the maximum number
+        # of samples under this assignment across workers.
+        window_indexes = range(len(self.windows))
+        windows_by_worker = [
+            window_indexes[cur_rank :: self.world_size][cur_worker_id::num_workers]
+            for cur_rank in range(self.world_size)
+            for cur_worker_id in range(num_workers)
+        ]
+
+        # Now compute the maximum number of samples across workers.
+        max_num_patches = 0
+        for worker_windows in windows_by_worker:
+            worker_num_patches = 0
+            for window_id in worker_windows:
+                worker_num_patches += self.get_window_num_patches(
+                    self.windows[window_id].bounds
+                )
+            max_num_patches = max(max_num_patches, worker_num_patches)
+
+        # Each worker needs at least one window, otherwise it won't be able to pad.
+        # Unless there are zero windows total, which is fine.
+        # Previously we would address this by borrowing the windows from another
+        # worker, but this causes issues with RslearnWriter: if we yield the same
+        # window from parallel workers, it may end up writing an empty output for that
+        # window in the end.
+        # So now we raise an error instead, and require the number of workers to be
+        # less than the number of windows.
+        if len(windows_by_worker[global_worker_id]) == 0 and max_num_patches > 0:
+            raise ValueError(
+                f"the number of workers {global_num_workers} must be <= the number of windows {len(self.windows)}"
+            )
+
+        return (windows_by_worker[global_worker_id], max_num_patches)
+
+    def __iter__(
+        self,
+    ) -> Iterator[tuple[dict[str, Any], dict[str, Any], dict[str, Any]]]:
+        """Iterate over all patches in each element of the underlying ModelDataset."""
+        # Iterate over the window IDs until we have returned enough samples.
+        window_ids, num_samples_needed = self._get_worker_iteration_data()
+        num_samples_returned = 0
+
+        for iteration_idx in itertools.count():
+            for window_id in window_ids:
+                raw_inputs, passthrough_inputs, metadata = self.dataset.get_raw_inputs(
+                    window_id
+                )
+                bounds = metadata["bounds"]
+
+                # For simplicity, pad tensors by patch size to ensure that any patch bounds
+                # extending outside the window bounds will not have issues when we slice
+                # the tensors later.
+                pad_slice_protect(raw_inputs, passthrough_inputs, self.patch_size)
+
+                # Now iterate over the patches and extract/yield the crops.
+                # Note that, in case user is leveraging RslearnWriter, it is important that
+                # the patch_idx be increasing (as we iterate) within one window.
+                patches = get_window_patch_options(
+                    self.patch_size, self.overlap_size, bounds
+                )
+                for patch_idx, patch_bounds in enumerate(patches):
+                    cur_geom = STGeometry(
+                        metadata["projection"], shapely.box(*patch_bounds), None
+                    )
+                    start_offset = (
+                        patch_bounds[0] - bounds[0],
+                        patch_bounds[1] - bounds[1],
+                    )
+                    end_offset = (
+                        patch_bounds[2] - bounds[0],
+                        patch_bounds[3] - bounds[1],
+                    )
+
+                    # Define a helper function to handle each input dict.
+                    def crop_input_dict(d: dict[str, Any]) -> dict[str, Any]:
+                        cropped = {}
+                        for input_name, value in d.items():
+                            if isinstance(value, torch.Tensor):
+                                # Crop the CHW tensor.
+                                cropped[input_name] = value[
+                                    :,
+                                    start_offset[1] : end_offset[1],
+                                    start_offset[0] : end_offset[0],
+                                ].clone()
+                            elif isinstance(value, list):
+                                cropped[input_name] = [
+                                    feat
+                                    for feat in value
+                                    if cur_geom.intersects(feat.geometry)
+                                ]
+                            else:
+                                raise ValueError(
+                                    "got input that is neither tensor nor feature list"
+                                )
+                        return cropped
+
+                    cur_raw_inputs = crop_input_dict(raw_inputs)
+                    cur_passthrough_inputs = crop_input_dict(passthrough_inputs)
+
+                    # Adjust the metadata as well.
+                    cur_metadata = metadata.copy()
+                    cur_metadata["bounds"] = patch_bounds
+                    cur_metadata["patch_idx"] = patch_idx
+                    cur_metadata["num_patches"] = len(patches)
+
+                    # Now we can compute input and target dicts via the task.
+                    input_dict, target_dict = self.dataset.task.process_inputs(
+                        cur_raw_inputs,
+                        metadata=cur_metadata,
+                        load_targets=not self.dataset.split_config.get_skip_targets(),
+                    )
+                    input_dict.update(cur_passthrough_inputs)
+                    input_dict, target_dict = self.dataset.transforms(
+                        input_dict, target_dict
+                    )
+                    input_dict["dataset_source"] = self.dataset.name
+
+                    if num_samples_returned < num_samples_needed:
+                        yield input_dict, target_dict, cur_metadata
+                        num_samples_returned += 1
+                    else:
+                        assert iteration_idx > 0
+
+            if num_samples_returned >= num_samples_needed:
+                break
+
+    def get_dataset_examples(self) -> list[Window]:
+        """Returns a list of windows in this dataset."""
+        return self.dataset.get_dataset_examples()
+
+
+class InMemoryAllPatchesDataset(torch.utils.data.Dataset):
+    """This wraps a ModelDataset to iterate over all patches in that dataset.
+
+    This should be used when SplitConfig.load_all_patches is enabled.
+
+    This is a simpler version of IterableAllPatchesDataset that caches all windows in memory.
+    This is useful for small datasets that fit in memory.
+    """
+
+    def __init__(
+        self,
+        dataset: ModelDataset,
+        patch_size: tuple[int, int],
+        overlap_ratio: float = 0.0,
+    ):
+        """Create a new InMemoryAllPatchesDataset.
+
+        Args:
+            dataset: the ModelDataset to wrap.
+            patch_size: the size of the patches to extract.
+            overlap_ratio: whether to include overlap between the patches. Note that
+                the right/bottom-most patches may still overlap since we ensure that
+                all patches are contained in the window bounds.
+        """
+        super().__init__()
+        self.dataset = dataset
+        self.patch_size = patch_size
+        self.overlap_size = (
+            round(self.patch_size[0] * overlap_ratio),
+            round(self.patch_size[1] * overlap_ratio),
+        )
+        self.windows = self.dataset.get_dataset_examples()
+        self.window_cache: dict[
+            int, tuple[dict[str, Any], dict[str, Any], dict[str, Any]]
+        ] = {}
+
+        # Precompute the batch boundaries for each window
+        self.patches = []
+        for window_id, window in enumerate(self.windows):
+            patch_bounds = get_window_patch_options(
+                self.patch_size, self.overlap_size, window.bounds
+            )
+            for i, patch_bound in enumerate(patch_bounds):
+                self.patches.append((window_id, patch_bound, (i, len(patch_bounds))))
+
+    def get_raw_inputs(
+        self, index: int
+    ) -> tuple[dict[str, Any], dict[str, Any], dict[str, Any]]:
+        """Get the raw inputs for a single patch. Retrieve from cache if possible.
+
+        Also crops/pads the tensors by patch size to protect slicing near right/bottom edges.
+
+        Args:
+            index: the index of the patch.
+
+        Returns:
+            a tuple of (raw_inputs, passthrough_inputs, metadata).
+        """
+        if index in self.window_cache:
+            return self.window_cache[index]
+
+        raw_inputs, passthrough_inputs, metadata = self.dataset.get_raw_inputs(index)
+        pad_slice_protect(raw_inputs, passthrough_inputs, self.patch_size)
+
+        self.window_cache[index] = (raw_inputs, passthrough_inputs, metadata)
+        return self.window_cache[index]
+
+    @staticmethod
+    def _crop_input_dict(
+        d: dict[str, Any],
+        start_offset: tuple[int, int],
+        end_offset: tuple[int, int],
+        cur_geom: STGeometry,
+    ) -> dict[str, Any]:
+        """Crop a dictionary of inputs to the given bounds."""
+        cropped = {}
+        for input_name, value in d.items():
+            if isinstance(value, torch.Tensor):
+                cropped[input_name] = value[
+                    :,
+                    start_offset[1] : end_offset[1],
+                    start_offset[0] : end_offset[0],
+                ].clone()
+            elif isinstance(value, list):
+                cropped[input_name] = [
+                    feat for feat in value if cur_geom.intersects(feat.geometry)
+                ]
+            else:
+                raise ValueError("got input that is neither tensor nor feature list")
+        return cropped
+
+    def __len__(self) -> int:
+        """Return the total number of patches in the dataset."""
+        return len(self.patches)
+
+    def __getitem__(
+        self, index: int
+    ) -> tuple[dict[str, Any], dict[str, Any], dict[str, Any]]:
+        """Return (input_dict, target_dict, metadata) for a single flattened patch."""
+        (window_id, patch_bounds, (patch_idx, num_patches)) = self.patches[index]
+        raw_inputs, passthrough_inputs, metadata = self.get_raw_inputs(window_id)
+        bounds = metadata["bounds"]
+
+        cur_geom = STGeometry(metadata["projection"], shapely.box(*patch_bounds), None)
+        start_offset = (patch_bounds[0] - bounds[0], patch_bounds[1] - bounds[1])
+        end_offset = (patch_bounds[2] - bounds[0], patch_bounds[3] - bounds[1])
+
+        cur_raw_inputs = self._crop_input_dict(
+            raw_inputs, start_offset, end_offset, cur_geom
+        )
+        cur_passthrough_inputs = self._crop_input_dict(
+            passthrough_inputs, start_offset, end_offset, cur_geom
+        )
+
+        # Adjust the metadata as well.
+        cur_metadata = metadata.copy()
+        cur_metadata["bounds"] = patch_bounds
+        cur_metadata["patch_idx"] = patch_idx
+        cur_metadata["num_patches"] = num_patches
+
+        # Now we can compute input and target dicts via the task.
+        input_dict, target_dict = self.dataset.task.process_inputs(
+            cur_raw_inputs,
+            metadata=cur_metadata,
+            load_targets=not self.dataset.split_config.get_skip_targets(),
+        )
+        input_dict.update(cur_passthrough_inputs)
+        input_dict, target_dict = self.dataset.transforms(input_dict, target_dict)
+        input_dict["dataset_source"] = self.dataset.name
+
+        return input_dict, target_dict, cur_metadata
+
+    def get_dataset_examples(self) -> list[Window]:
+        """Returns a list of windows in this dataset."""
+        return self.dataset.get_dataset_examples()
+
+    def set_name(self, name: str) -> None:
+        """Sets dataset name.
+
+        Args:
+            name: dataset name
+        """
+        self.dataset.set_name(name)

rslearn/train/data_module.py

@@ -15,10 +15,12 @@ from rslearn.dataset import Dataset
 from rslearn.log_utils import get_logger
 from rslearn.train.tasks import Task
 
-from .dataset import (
-    DataInput,
+from .all_patches_dataset import (
     InMemoryAllPatchesDataset,
     IterableAllPatchesDataset,
+)
+from .dataset import (
+    DataInput,
     ModelDataset,
     MultiDataset,
     RetryDataset,