ObjectDetectionGeoDataConfig leaves window_opts empty when created with scene_dataset

[aerotractjack]

I am attempting to use the ObjectDetectionLearner, and creating my data from a collection of scenes (I have multiple scenes for training and validation). When I attempt to use a collection of scenes within a DatasetConfig, I get an error from the learner constructor saying my window_opts do not contain keys corresponding to the scene name. When I examine the window_opts, they are empty. I must be missing something but I can't find help in the docs. Here is my relevant code to create the scenes and learner. Any ideas? Thank you so much

    def make_scene(self, name, image_uri, label_uri, aoi_uri):
        '''' Define a Scene with image and labels from the given URIs. '''
        raster_source = RasterioSourceConfig(
                uris=[image_uri], channel_order=[0, 1, 2])
        label_source = ObjectDetectionLabelSourceConfig(
            vector_source=GeoJSONVectorSourceConfig(
                    uri=label_uri,
                    ignore_crs_field=True,
                    transformers=[ClassInferenceTransformerConfig(default_class_id=0)]
            )
        )
        aoi_uris = [aoi_uri] if aoi_uri is not None else []
        sconf = SceneConfig(
            id=name,
            raster_source=raster_source,
            label_source=label_source,
            aoi_uris=aoi_uris
        )
        return sconf
    
    def _build_scene_ds(self, uris):
        scenes = []
        for i in range(len(uris["image"])):
            s = self.make_scene(
                uris["name"][i], uris["image"][i], uris["labels"][i], uris["aoi"][i]
            )
            scenes.append(s)
        return scenes
    
    def build_train_val_scene_ds(self):
        train = self._build_scene_ds(self.train_uris)
        val = self._build_scene_ds(self.val_uris)
        dconf = DatasetConfig(
            class_config=self.cc, train_scenes=train, validation_scenes=val)
        return dconf

    def configure_training(self):
        data_cfg = ObjectDetectionGeoDataConfig(
            class_names=self.cc.names,
            class_colors=self.cc.colors,
            scene_dataset=self.build_train_val_scene_ds(),
            num_workers=2,
        )
        print(data_cfg.window_opts)  # --> `{}`
        solver_cfg = SolverConfig(
            batch_sz=4,
            lr=1e-3,
        )
        model = self.build_model()
        learner_cfg = ObjectDetectionLearnerConfig(
            data=data_cfg, 
            solver=solver_cfg,
            model=model)
        learner = ObjectDetectionLearner(  # ERROR HERE
            cfg=learner_cfg,
            output_dir="./demo",
            # train_ds=self.build_train_ds(),
            # valid_ds=self.build_val_ds()
        )
        return learner

Results in this error:

Traceback (most recent call last):
  File "/home/aerotract/software/rvml-pipeline/train.py", line 212, in <module>
    util.run_if_main(__name__, RVBase)
  File "/home/aerotract/software/rvml-pipeline/util.py", line 18, in run_if_main
    obj.run()
  File "/home/aerotract/software/rvml-pipeline/train.py", line 210, in run
    self.train()
  File "/home/aerotract/software/rvml-pipeline/train.py", line 205, in train
    learner = self.configure_training()
  File "/home/aerotract/software/rvml-pipeline/train.py", line 196, in configure_training
    learner = ObjectDetectionLearner(
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/learner.py", line 239, in __init__
    self.setup_training(loss_def_path=loss_def_path)
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/learner.py", line 289, in setup_training
    self.setup_data()
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/learner.py", line 413, in setup_data
    train_ds, valid_ds, test_ds = self.build_datasets()
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/learner.py", line 425, in build_datasets
    train_ds, val_ds, test_ds = self.cfg.data.build(
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/learner_config.py", line 1271, in build
    train_ds, val_ds, test_ds = self.make_datasets(
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/learner_config.py", line 1235, in make_datasets
    train_ds_list = [
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/learner_config.py", line 1236, in <listcomp>
    self.scene_to_dataset(s, train_tf, **kwargs) for s in train_scenes
  File "/home/aerotract/miniconda3/envs/ml/lib/python3.9/site-packages/rastervision/pytorch_learner/object_detection_learner_config.py", line 98, in scene_to_dataset
    opts = self.window_opts[scene.id]
KeyError: 'scene-key-name'

Thank you

[aerotractjack]

I was forgetting to set my chip and window options, the following fixed it for me:

def configure_training(self):
        chip_options = ObjectDetectionChipOptions(**self.kw.get("chip_kw", {}))
        chip_sz = self.kw["window_kw"].get("chip_sz", 450)
        window_opts = ObjectDetectionGeoDataWindowConfig(
            method=GeoDataWindowMethod.random,
            size=chip_sz,
            size_lims=(chip_sz, chip_sz + 1),
            max_windows=self.kw["window_kw"].get("max_windows", None),
            clip=True,
            neg_ratio=chip_options.neg_ratio,
            ioa_thresh=chip_options.ioa_thresh
        )
        data_cfg = ObjectDetectionGeoDataConfig(
            class_names=self.cc.names,
            class_colors=self.cc.colors,
            scene_dataset=self.build_train_val_scene_ds(),
            window_opts=window_opts,
            num_workers=2,
        )
        solver_cfg = SolverConfig(
            batch_sz=4,
            lr=1e-3,
        )
        model = self.build_model()
        learner_cfg = ObjectDetectionLearnerConfig(
            data=data_cfg, 
            solver=solver_cfg,
            model=model)
        learner = ObjectDetectionLearner(
            cfg=learner_cfg,
            output_dir="./demo",
            # train_ds=self.build_train_ds(),
            # valid_ds=self.build_val_ds()
        )
        return learner