lazy all_train_cameras

Summary: Avoid calculating all_train_cameras before it is needed, because it is slow in some datasets.

Reviewed By: shapovalov

Differential Revision: D38037157

fbshipit-source-id: 95461226655cde2626b680661951ab17ebb0ec75

projects/implicitron_trainer/experiment.py

@@ -391,7 +391,6 @@ def run_training(cfg: DictConfig) -> None:
     datasource = ImplicitronDataSource(**cfg.data_source_args)
     datasets, dataloaders = datasource.get_datasets_and_dataloaders()
     task = datasource.get_task()
-    all_train_cameras = datasource.get_all_train_cameras()
 
     # init the model
     model, stats, optimizer_state = init_model(cfg=cfg, accelerator=accelerator)
@@ -405,7 +404,7 @@ def run_training(cfg: DictConfig) -> None:
         _eval_and_dump(
             cfg,
             task,
-            all_train_cameras,
+            datasource.all_train_cameras,
             datasets,
             dataloaders,
             model,
@@ -490,7 +489,7 @@ def run_training(cfg: DictConfig) -> None:
             ):
                 _run_eval(
                     model,
-                    all_train_cameras,
+                    datasource.all_train_cameras,
                     dataloaders.test,
                     task,
                     camera_difficulty_bin_breaks=cfg.camera_difficulty_bin_breaks,
@@ -525,7 +524,7 @@ def run_training(cfg: DictConfig) -> None:
         _eval_and_dump(
             cfg,
             task,
-            all_train_cameras,
+            datasource.all_train_cameras,
             datasets,
             dataloaders,
             model,

pytorch3d/implicitron/dataset/data_source.py

@@ -30,9 +30,10 @@ class DataSourceBase(ReplaceableBase):
     def get_datasets_and_dataloaders(self) -> Tuple[DatasetMap, DataLoaderMap]:
         raise NotImplementedError()
 
-    def get_all_train_cameras(self) -> Optional[CamerasBase]:
+    @property
+    def all_train_cameras(self) -> Optional[CamerasBase]:
         """
-        If the data is all for a single scene, returns a list
+        If the data is all for a single scene, a list
         of the known training cameras for that scene, which is
         used for evaluating the viewpoint difficulty of the
         unseen cameras.
@@ -59,6 +60,7 @@ class ImplicitronDataSource(DataSourceBase):  # pyre-ignore[13]
 
     def __post_init__(self):
         run_auto_creation(self)
+        self._all_train_cameras_cache: Optional[Tuple[Optional[CamerasBase]]] = None
 
     def get_datasets_and_dataloaders(self) -> Tuple[DatasetMap, DataLoaderMap]:
         datasets = self.dataset_map_provider.get_dataset_map()
@@ -68,5 +70,10 @@ class ImplicitronDataSource(DataSourceBase):  # pyre-ignore[13]
     def get_task(self) -> Task:
         return self.dataset_map_provider.get_task()
 
-    def get_all_train_cameras(self) -> Optional[CamerasBase]:
+    @property
-        return self.dataset_map_provider.get_all_train_cameras()
+    def all_train_cameras(self) -> Optional[CamerasBase]:
+        if self._all_train_cameras_cache is None:  # pyre-ignore[16]
+            all_train_cameras = self.dataset_map_provider.get_all_train_cameras()
+            self._all_train_cameras_cache = (all_train_cameras,)
+
+        return self._all_train_cameras_cache[0]

pytorch3d/implicitron/eval_demo.py

@@ -118,8 +118,6 @@ def evaluate_dbir_for_category(
     if test_dataset is None or test_dataloader is None:
         raise ValueError("must have a test dataset.")
 
-    all_train_cameras = data_source.get_all_train_cameras()
-
     image_size = cast(JsonIndexDataset, test_dataset).image_width
 
     if image_size is None:
@@ -149,7 +147,7 @@ def evaluate_dbir_for_category(
                 preds["implicitron_render"],
                 bg_color=bg_color,
                 lpips_model=lpips_model,
-                source_cameras=all_train_cameras,
+                source_cameras=data_source.all_train_cameras,
             )
         )
 

tests/implicitron/test_batch_sampler.py

@@ -10,6 +10,7 @@ from collections import defaultdict
 from dataclasses import dataclass
 from itertools import product
 
+import numpy as np
 from pytorch3d.implicitron.dataset.data_loader_map_provider import (
     DoublePoolBatchSampler,
 )
@@ -53,6 +54,7 @@ class MockDataset(DatasetBase):
 
 class TestSceneBatchSampler(unittest.TestCase):
     def setUp(self):
+        np.random.seed(42)
         self.dataset_overfit = MockDataset(1)
 
     def test_overfit(self):

tests/implicitron/test_data_json_index.py

@@ -31,7 +31,7 @@ class TestDataJsonIndex(TestCaseMixin, unittest.TestCase):
 
         data_source = ImplicitronDataSource(**args)
 
-        cameras = data_source.get_all_train_cameras()
+        cameras = data_source.all_train_cameras
         self.assertIsInstance(cameras, PerspectiveCameras)
         self.assertEqual(len(cameras), 81)
 

tests/implicitron/test_data_llff.py

@@ -152,6 +152,6 @@ class TestDataLlff(TestCaseMixin, unittest.TestCase):
             self.assertEqual(i.frame_type, ["unseen"])
             self.assertEqual(i.image_rgb.shape, (1, 3, 800, 800))
 
-        cameras = data_source.get_all_train_cameras()
+        cameras = data_source.all_train_cameras
         self.assertIsInstance(cameras, PerspectiveCameras)
         self.assertEqual(len(cameras), 100)