Overhaul Trainers:

- consolidate `Trainer` and `TrainerConfig` classes
- move snapshots/num_workers/batch_size etc. to `RunConfig`
- make interface a simple `torch.utils.data.IterableDataset`.
- update `train.py` and `validate.py` to use overhauled `Trainer` classes

dacapo/experiments/trainers/__init__.py

@@ -1,5 +1,5 @@
 from .trainer import Trainer  # noqa
 from .trainer_config import TrainerConfig  # noqa
-from .dummy_trainer_config import DummyTrainerConfig, DummyTrainer  # noqa
-from .gunpowder_trainer_config import GunpowderTrainerConfig, GunpowderTrainer  # noqa
+from .dummy_trainer_config import DummyTrainerConfig  # noqa
+from .gunpowder_trainer_config import GunpowderTrainerConfig  # noqa
 from .gp_augments import AugmentConfig  # noqa

dacapo/experiments/trainers/dummy_trainer_config.py

@@ -1,10 +1,30 @@
 import attr
 
-from .dummy_trainer import DummyTrainer
+from dacapo.experiments.datasplits.datasets import DatasetConfig
+from dacapo.experiments.tasks.predictors import Predictor
 from .trainer_config import TrainerConfig
 
+from funlib.geometry import Roi, Coordinate
+
 from typing import Tuple
 
+import numpy as np
+import torch
+
+
+class GeneratorDataset(torch.utils.data.IterableDataset):
+    """
+    Helper class to return a torch IterableDataset from a generator
+    """
+
+    def __init__(self, generator, *args, **kwargs):
+        self.generator = generator
+        self.args = args
+        self.kwargs = kwargs
+
+    def __iter__(self):
+        return self.generator(*self.args, **self.kwargs)
+
 
 @attr.s
 class DummyTrainerConfig(TrainerConfig):
@@ -21,18 +41,31 @@ class DummyTrainerConfig(TrainerConfig):
 
     """
 
-    trainer_type = DummyTrainer
+    dummy_attr: bool = attr.ib(metadata={"help_text": "Dummy attribute."})
 
-    mirror_augment: bool = attr.ib(metadata={"help_text": "Dummy attribute."})
+    def iterable_dataset(
+        self,
+        datasets: list[DatasetConfig],
+        input_shape: Coordinate,
+        output_shape: Coordinate,
+        predictor: Predictor | None = None,
+    ):
+        in_roi = Roi(input_shape * 0, input_shape)
+        out_roi = Roi(output_shape * 0, output_shape)
+        in_voxel_size = datasets[0].raw.voxel_size
+        raw = torch.from_numpy(
+            datasets[0].raw[in_roi * in_voxel_size].astype(np.float32)
+        )
+        out_raw = torch.from_numpy(
+            datasets[0].raw[out_roi * in_voxel_size].astype(np.float32)
+        )
 
-    def verify(self) -> Tuple[bool, str]:
-        """
-        Verify the DummyTrainerConfig object.
+        def generator():
+            while True:
+                yield {
+                    "raw": raw,
+                    "target": out_raw,
+                    "weight": torch.ones_like(out_raw),
+                }
 
-        Returns:
-            Tuple[bool, str]: A tuple containing a boolean value indicating whether the DummyTrainerConfig object is valid
-                and a string containing the reason why the object is invalid.
-        Examples:
-            >>> valid, reason = trainer_config.verify()
-        """
-        return False, "This is a DummyTrainerConfig and is never valid"
+        return GeneratorDataset(generator)