rslearn 0.0.11__py3-none-any.whl → 0.0.13__py3-none-any.whl

rslearn/config/dataset.py

@@ -125,7 +125,8 @@ class BandSetConfig:
         self,
         config_dict: dict[str, Any],
         dtype: DType,
-        bands: list[str],
+        bands: list[str] | None = None,
+        num_bands: int | None = None,
         format: dict[str, Any] | None = None,
         zoom_offset: int = 0,
         remap: dict[str, Any] | None = None,
@@ -137,7 +138,10 @@ class BandSetConfig:
         Args:
             config_dict: the config dict used to configure this BandSetConfig
             dtype: the pixel value type to store tiles in
-            bands: list of band names in this BandSetConfig
+            bands: list of band names in this BandSetConfig. One of bands or num_bands
+                must be set.
+            num_bands: the number of bands in this band set. The bands will be named
+                B00, B01, B02, etc.
             format: the format to store tiles in, defaults to geotiff
             zoom_offset: store images at a resolution higher or lower than the window
                 resolution. This enables keeping source data at its native resolution,
@@ -155,6 +159,14 @@ class BandSetConfig:
                 materialization when creating mosaics, to determine which parts of the
                 source images should be copied.
         """
+        if (bands is None and num_bands is None) or (
+            bands is not None and num_bands is not None
+        ):
+            raise ValueError("exactly one of bands and num_bands must be set")
+        if bands is None:
+            assert num_bands is not None
+            bands = [f"B{idx}" for idx in range(num_bands)]
+
         if class_names is not None and len(bands) != len(class_names):
             raise ValueError(
                 f"the number of class lists ({len(class_names)}) does not match the number of bands ({len(bands)})"
@@ -187,9 +199,16 @@ class BandSetConfig:
         kwargs = dict(
             config_dict=config,
             dtype=DType(config["dtype"]),
-            bands=config["bands"],
         )
-        for k in ["format", "zoom_offset", "remap", "class_names", "nodata_vals"]:
+        for k in [
+            "bands",
+            "num_bands",
+            "format",
+            "zoom_offset",
+            "remap",
+            "class_names",
+            "nodata_vals",
+        ]:
             if k in config:
                 kwargs[k] = config[k]
         return BandSetConfig(**kwargs)  # type: ignore

rslearn/data_sources/planetary_computer.py

@@ -827,3 +827,55 @@ class Sentinel1(PlanetaryComputer):
                 kwargs[k] = d[k]
 
         return Sentinel1(**kwargs)
+
+
+class Naip(PlanetaryComputer):
+    """A data source for NAIP data on Microsoft Planetary Computer.
+
+    See https://planetarycomputer.microsoft.com/dataset/naip.
+    """
+
+    COLLECTION_NAME = "naip"
+    ASSET_BANDS = {"image": ["R", "G", "B", "NIR"]}
+
+    def __init__(
+        self,
+        **kwargs: Any,
+    ):
+        """Initialize a new Naip instance.
+
+        Args:
+            band_names: list of bands to try to ingest.
+            kwargs: additional arguments to pass to PlanetaryComputer.
+        """
+        super().__init__(
+            collection_name=self.COLLECTION_NAME,
+            asset_bands=self.ASSET_BANDS,
+            **kwargs,
+        )
+
+    @staticmethod
+    def from_config(config: RasterLayerConfig, ds_path: UPath) -> "Naip":
+        """Creates a new Naip instance from a configuration dictionary."""
+        if config.data_source is None:
+            raise ValueError("config.data_source is required")
+        d = config.data_source.config_dict
+        kwargs = {}
+
+        if "timeout_seconds" in d:
+            kwargs["timeout"] = timedelta(seconds=d["timeout_seconds"])
+
+        if "cache_dir" in d:
+            kwargs["cache_dir"] = join_upath(ds_path, d["cache_dir"])
+
+        simple_optionals = [
+            "query",
+            "sort_by",
+            "sort_ascending",
+            "max_items_per_client",
+        ]
+        for k in simple_optionals:
+            if k in d:
+                kwargs[k] = d[k]
+
+        return Naip(**kwargs)

rslearn/dataset/handler_summaries.py

@@ -20,6 +20,7 @@ class LayerPrepareSummary:
     # Counts
     windows_prepared: int
     windows_skipped: int
+    windows_rejected: int
     get_items_attempts: int
 
 

rslearn/dataset/manage.py

@@ -118,6 +118,7 @@ def prepare_dataset_windows(
                     duration_seconds=time.monotonic() - layer_start_time,
                     windows_prepared=0,
                     windows_skipped=len(windows),
+                    windows_rejected=0,
                     get_items_attempts=0,
                 )
             )
@@ -141,6 +142,7 @@ def prepare_dataset_windows(
                     duration_seconds=time.monotonic() - layer_start_time,
                     windows_prepared=0,
                     windows_skipped=len(windows),
+                    windows_rejected=0,
                     get_items_attempts=0,
                 )
             )
@@ -181,6 +183,9 @@ def prepare_dataset_windows(
             attempts_counter=attempts_counter,
         )
 
+        windows_prepared = 0
+        windows_rejected = 0
+        min_matches = data_source_cfg.query_config.min_matches
         for window, result in zip(needed_windows, results):
             layer_datas = window.load_layer_datas()
             layer_datas[layer_name] = WindowLayerData(
@@ -191,13 +196,22 @@ def prepare_dataset_windows(
             )
             window.save_layer_datas(layer_datas)
 
+            # If result is empty and min_matches > 0, window was rejected due to min_matches
+            if len(result) == 0 and min_matches > 0:
+                windows_rejected += 1
+            else:
+                windows_prepared += 1
+
+        windows_skipped = len(windows) - len(needed_windows)
+
         layer_summaries.append(
             LayerPrepareSummary(
                 layer_name=layer_name,
                 data_source_name=data_source_cfg.name,
                 duration_seconds=time.monotonic() - layer_start_time,
-                windows_prepared=len(needed_windows),  # we assume all have succeeded
-                windows_skipped=len(windows) - len(needed_windows),
+                windows_prepared=windows_prepared,
+                windows_skipped=windows_skipped,
+                windows_rejected=windows_rejected,
                 get_items_attempts=attempts_counter.value,
             )
         )

rslearn/models/anysat.py

@@ -1,4 +1,8 @@
-"""AnySat model."""
+"""AnySat model.
+
+This code loads the AnySat model from torch hub. See
+https://github.com/gastruc/AnySat for applicable license and copyright information.
+"""
 
 from typing import Any
 

rslearn/models/dinov3.py

@@ -1,4 +1,9 @@
-"""DinoV3 model."""
+"""DinoV3 model.
+
+This code loads the DINOv3 model. You must obtain the model separately from Meta to use
+it. See https://github.com/facebookresearch/dinov3 for applicable license and copyright
+information.
+"""
 
 from enum import StrEnum
 from pathlib import Path

rslearn/models/feature_center_crop.py

@@ -0,0 +1,50 @@
+"""Apply center cropping on a feature map."""
+
+from typing import Any
+
+import torch
+
+
+class FeatureCenterCrop(torch.nn.Module):
+    """Apply center cropping on the input feature maps."""
+
+    def __init__(
+        self,
+        sizes: list[tuple[int, int]],
+    ) -> None:
+        """Create a new FeatureCenterCrop.
+
+        Only the center of each feature map will be retained and passed to the next
+        module.
+
+        Args:
+            sizes: a list of (height, width) tuples, with one tuple for each input
+                feature map.
+        """
+        super().__init__()
+        self.sizes = sizes
+
+    def forward(
+        self, features: list[torch.Tensor], inputs: list[dict[str, Any]]
+    ) -> list[torch.Tensor]:
+        """Apply center cropping on the feature maps.
+
+        Args:
+            features: list of feature maps at different resolutions.
+            inputs: original inputs (ignored).
+
+        Returns:
+            center cropped feature maps.
+        """
+        new_features = []
+        for i, feat in enumerate(features):
+            height, width = self.sizes[i]
+            if feat.shape[2] < height or feat.shape[3] < width:
+                raise ValueError(
+                    "feature map is smaller than the desired height and width"
+                )
+            start_h = feat.shape[2] // 2 - height // 2
+            start_w = feat.shape[3] // 2 - width // 2
+            feat = feat[:, :, start_h : start_h + height, start_w : start_w + width]
+            new_features.append(feat)
+        return new_features

rslearn/models/olmoearth_pretrain/model.py

@@ -9,6 +9,11 @@ from einops import rearrange
 from olmo_core.config import Config
 from olmo_core.distributed.checkpoint import load_model_and_optim_state
 from olmoearth_pretrain.data.constants import Modality
+from olmoearth_pretrain.model_loader import (
+    ModelID,
+    load_model_from_id,
+    load_model_from_path,
+)
 from olmoearth_pretrain.nn.flexihelios import Encoder, TokensAndMasks
 from olmoearth_pretrain.train.masking import MaskedOlmoEarthSample, MaskValue
 from upath import UPath
@@ -31,54 +36,115 @@ AUTOCAST_DTYPE_MAP = {
     "float32": torch.float32,
 }
 
+EMBEDDING_SIZES = {
+    ModelID.OLMOEARTH_V1_NANO: 128,
+    ModelID.OLMOEARTH_V1_TINY: 192,
+    ModelID.OLMOEARTH_V1_BASE: 768,
+    ModelID.OLMOEARTH_V1_LARGE: 1024,
+}
+
 
 class OlmoEarth(torch.nn.Module):
     """A wrapper to support the OlmoEarth model."""
 
     def __init__(
         self,
-        # TODO: we should accept model ID instead of checkpoint_path once we are closer
-        # to being ready for release.
-        checkpoint_path: str,
-        selector: list[str | int] = [],
+        patch_size: int,
+        model_id: ModelID | None = None,
+        model_path: str | None = None,
+        checkpoint_path: str | None = None,
+        selector: list[str | int] = ["encoder"],
         forward_kwargs: dict[str, Any] = {},
         random_initialization: bool = False,
         embedding_size: int | None = None,
-        patch_size: int | None = None,
         autocast_dtype: str | None = "bfloat16",
     ):
         """Create a new OlmoEarth model.
 
         Args:
-            checkpoint_path: the checkpoint directory to load. It should contain
-                config.json file as well as model_and_optim folder.
+            patch_size: token spatial patch size to use.
+            model_id: the model ID to load. One of model_id or model_path or checkpoint_path must be
+                set.
+            model_path: the path to load the model from. One of model_id or model_path or checkpoint_path must be
+                set. Same structure as the HF-hosted `model_id` models: bundle with a config.json and weights.pth.
+            checkpoint_path: the checkpoint directory to load from, if model_id or model_path is not
+                set. It should contain a distributed checkpoint with a config.json file as well as model_and_optim
+                folder.
             selector: an optional sequence of attribute names or list indices to select
-                the sub-module that should be applied on the input images.
+                the sub-module that should be applied on the input images. Defaults to
+                ["encoder"] to select only the transformer encoder.
             forward_kwargs: additional arguments to pass to forward pass besides the
                  MaskedOlmoEarthSample.
             random_initialization: whether to skip loading the checkpoint so the
                 weights are randomly initialized. In this case, the checkpoint is only
                 used to define the model architecture.
             embedding_size: optional embedding size to report via
-                get_backbone_channels.
-            patch_size: optional patch size to report via get_backbone_channels.
+                get_backbone_channels (if model_id is not set).
             autocast_dtype: which dtype to use for autocasting, or set None to disable.
         """
+        if (
+            sum(
+                [
+                    model_id is not None,
+                    model_path is not None,
+                    checkpoint_path is not None,
+                ]
+            )
+            != 1
+        ):
+            raise ValueError(
+                "exactly one of model_id, model_path, or checkpoint_path must be set"
+            )
+
         super().__init__()
-        _checkpoint_path = UPath(checkpoint_path)
+        self.patch_size = patch_size
         self.forward_kwargs = forward_kwargs
         self.embedding_size = embedding_size
-        self.patch_size = patch_size
 
         if autocast_dtype is not None:
             self.autocast_dtype = AUTOCAST_DTYPE_MAP[autocast_dtype]
         else:
             self.autocast_dtype = None
 
+        if model_id is not None:
+            # Load from Hugging Face.
+            model = load_model_from_id(model_id, load_weights=not random_initialization)
+            if self.embedding_size is None and model_id in EMBEDDING_SIZES:
+                self.embedding_size = EMBEDDING_SIZES[model_id]
+
+        elif model_path is not None:
+            # Load from path.
+            model = load_model_from_path(
+                UPath(model_path), load_weights=not random_initialization
+            )
+
+        else:
+            # Load the distributed model checkpoint by path through Olmo Core
+            model = self._load_model_from_checkpoint(
+                UPath(checkpoint_path), random_initialization
+            )
+
+        # Select just the portion of the model that we actually want to use.
+        for part in selector:
+            if isinstance(part, str):
+                model = getattr(model, part)
+            else:
+                model = model[part]
+        self.model = model
+
+    def _load_model_from_checkpoint(
+        self, checkpoint_upath: UPath, random_initialization: bool
+    ) -> torch.nn.Module:
+        """Load the OlmoEarth pre-trained model from a distributed checkpoint folder.
+
+        The folder should contain config.json as well as the model_and_optim folder
+        that contains the distributed checkpoint. This is the format produced by
+        pre-training runs in olmoearth_pretrain.
+        """
         # Load the model config and initialize it.
         # We avoid loading the train module here because it depends on running within
         # olmo_core.
-        with (_checkpoint_path / "config.json").open() as f:
+        with (checkpoint_upath / "config.json").open() as f:
             config_dict = json.load(f)
             model_config = Config.from_dict(config_dict["model"])
 
@@ -86,22 +152,14 @@ class OlmoEarth(torch.nn.Module):
 
         # Load the checkpoint.
         if not random_initialization:
-            train_module_dir = _checkpoint_path / "model_and_optim"
+            train_module_dir = checkpoint_upath / "model_and_optim"
             if train_module_dir.exists():
                 load_model_and_optim_state(str(train_module_dir), model)
                 logger.info(f"loaded OlmoEarth encoder from {train_module_dir}")
             else:
                 logger.info(f"could not find OlmoEarth encoder at {train_module_dir}")
-        else:
-            logger.info("skipping loading OlmoEarth encoder")
 
-        # Select just the portion of the model that we actually want to use.
-        for part in selector:
-            if isinstance(part, str):
-                model = getattr(model, part)
-            else:
-                model = model[part]
-        self.model = model
+        return model
 
     def forward(self, inputs: list[dict[str, Any]]) -> list[torch.Tensor]:
         """Compute feature maps from the OlmoEarth backbone.
@@ -167,13 +225,16 @@ class OlmoEarth(torch.nn.Module):
             if isinstance(self.model, Encoder):
                 # Encoder has a fast_pass argument to indicate mask is not needed.
                 tokens_and_masks = self.model(
-                    sample, fast_pass=True, **self.forward_kwargs
+                    sample,
+                    fast_pass=True,
+                    patch_size=self.patch_size,
+                    **self.forward_kwargs,
                 )["tokens_and_masks"]
             else:
                 # Other models like STEncoder do not have this option supported.
-                tokens_and_masks = self.model(sample, **self.forward_kwargs)[
-                    "tokens_and_masks"
-                ]
+                tokens_and_masks = self.model(
+                    sample, patch_size=self.patch_size, **self.forward_kwargs
+                )["tokens_and_masks"]
 
         # Apply temporal/modality pooling so we just have one feature per patch.
         features = []

rslearn/models/prithvi.py

@@ -1,4 +1,12 @@
-"""Prithvi V2."""
+"""Prithvi V2.
+
+This code is adapted from https://github.com/NASA-IMPACT/Prithvi-WxC
+
+The code is released under:
+
+MIT License
+Copyright (c) 2024 Inter Agency Implementation and Advanced Concepts
+"""
 
 import json
 import logging

rslearn/train/lightning_module.py

@@ -94,7 +94,6 @@ class RslearnLightningModule(L.LightningModule):
         restore_config: RestoreConfig | None = None,
         print_parameters: bool = False,
         print_model: bool = False,
-        strict_loading: bool = True,
         # Deprecated options.
         lr: float = 1e-3,
         plateau: bool = False,
@@ -118,7 +117,6 @@ class RslearnLightningModule(L.LightningModule):
             print_parameters: whether to print the list of model parameters after model
                 initialization
             print_model: whether to print the model after model initialization
-            strict_loading: whether to strictly load the model parameters.
             lr: deprecated.
             plateau: deprecated.
             plateau_factor: deprecated.
@@ -132,7 +130,6 @@ class RslearnLightningModule(L.LightningModule):
         self.visualize_dir = visualize_dir
         self.metrics_file = metrics_file
         self.restore_config = restore_config
-        self.strict_loading = strict_loading
 
         self.scheduler_factory: SchedulerFactory | None = None
         if scheduler:

rslearn/train/prediction_writer.py

@@ -22,7 +22,11 @@ from rslearn.log_utils import get_logger
 from rslearn.utils.array import copy_spatial_array
 from rslearn.utils.feature import Feature
 from rslearn.utils.geometry import PixelBounds
-from rslearn.utils.raster_format import RasterFormat, load_raster_format
+from rslearn.utils.raster_format import (
+    RasterFormat,
+    adjust_projection_and_bounds_for_array,
+    load_raster_format,
+)
 from rslearn.utils.vector_format import VectorFormat, load_vector_format
 
 from .lightning_module import RslearnLightningModule
@@ -68,15 +72,18 @@ class VectorMerger(PatchPredictionMerger):
 class RasterMerger(PatchPredictionMerger):
     """Merger for raster data that copies the rasters to the output."""
 
-    def __init__(self, padding: int | None = None):
+    def __init__(self, padding: int | None = None, downsample_factor: int = 1):
         """Create a new RasterMerger.
 
         Args:
             padding: the padding around the individual patch outputs to remove. This is
                 typically used when leveraging overlapping patches. Portions of outputs
                 at the border of the window will still be retained.
+            downsample_factor: the factor by which the rasters output by the task are
+                lower in resolution relative to the window resolution.
         """
         self.padding = padding
+        self.downsample_factor = downsample_factor
 
     def merge(
         self, window: Window, outputs: Sequence[PendingPatchOutput]
@@ -87,8 +94,8 @@ class RasterMerger(PatchPredictionMerger):
         merged_image = np.zeros(
             (
                 num_channels,
-                window.bounds[3] - window.bounds[1],
-                window.bounds[2] - window.bounds[0],
+                (window.bounds[3] - window.bounds[1]) // self.downsample_factor,
+                (window.bounds[2] - window.bounds[0]) // self.downsample_factor,
             ),
             dtype=dtype,
         )
@@ -104,7 +111,10 @@ class RasterMerger(PatchPredictionMerger):
             # If the output is not on the left or top boundary, then we should apply
             # the padding (if set).
             src = output.output
-            src_offset = (output.bounds[0], output.bounds[1])
+            src_offset = (
+                output.bounds[0] // self.downsample_factor,
+                output.bounds[1] // self.downsample_factor,
+            )
             if self.padding is not None and output.bounds[0] != window.bounds[0]:
                 src = src[:, :, self.padding :]
                 src_offset = (src_offset[0] + self.padding, src_offset[1])
@@ -116,7 +126,10 @@ class RasterMerger(PatchPredictionMerger):
                 src=src,
                 dst=merged_image,
                 src_offset=src_offset,
-                dst_offset=(window.bounds[0], window.bounds[1]),
+                dst_offset=(
+                    window.bounds[0] // self.downsample_factor,
+                    window.bounds[1] // self.downsample_factor,
+                ),
             )
 
         return merged_image
@@ -330,9 +343,13 @@ class RslearnWriter(BasePredictionWriter):
                 self.output_layer, self.layer_config.band_sets[0].bands
             )
             assert isinstance(self.format, RasterFormat)
-            self.format.encode_raster(
-                raster_dir, window.projection, window.bounds, merged_output
+
+            # In case the merged_output is at a different resolution than the window,
+            # get adjusted projection and bounds for writing it.
+            projection, bounds = adjust_projection_and_bounds_for_array(
+                window.projection, window.bounds, merged_output
             )
+            self.format.encode_raster(raster_dir, projection, bounds, merged_output)
 
         elif self.layer_config.layer_type == LayerType.VECTOR:
             layer_dir = window.get_layer_dir(self.output_layer)

rslearn/train/tasks/classification.py

@@ -49,8 +49,8 @@ class ClassificationTask(BasicTask):
                 features with matching properties.
             read_class_id: whether to read an integer class ID instead of the class
                 name.
-            allow_invalid: instead of throwing error when no regression label is found
-                at a window, simply mark the example invalid for this task
+            allow_invalid: instead of throwing error when no classification label is
+                found at a window, simply mark the example invalid for this task
             skip_unknown_categories: whether to skip examples with categories that are
                 not passed via classes, instead of throwing error
             prob_property: when predicting, write probabilities in addition to class ID

rslearn/train/tasks/detection.py

@@ -72,11 +72,11 @@ class DetectionTask(BasicTask):
         f1_metric_kwargs: dict[str, Any] = {},
         **kwargs: Any,
     ) -> None:
-        """Initialize a new SegmentationTask.
+        """Initialize a new DetectionTask.
 
         Args:
-            property_name: the property from which to extract the class name. The class
-                is read from the first matching feature.
+            property_name: the property from which to extract the class name. Features
+                without this property name are ignored.
             classes: a list of class names.
             filters: optional list of (property_name, property_value) to only consider
                 features with matching properties.
@@ -86,8 +86,8 @@ class DetectionTask(BasicTask):
                 not passed via classes, instead of throwing error
             skip_empty_examples: whether to skip examples with zero labels.
             colors: optional colors for each class
-            box_size: force all boxes to be this size, centered at the centroid of the
-                geometry. Required for Point geometries.
+            box_size: force all boxes to be two times this size, centered at the
+                centroid of the geometry. Required for Point geometries.
             clip_boxes: whether to clip boxes to the image bounds.
             exclude_by_center: before optionally clipping boxes, exclude boxes if the
                 center is outside the image bounds.