Fix test evaluation for Blender data

Summary: Blender data doesn't have depths or crops.

Reviewed By: shapovalov

Differential Revision: D38345583

fbshipit-source-id: a19300daf666bbfd799d0038aeefa14641c559d7

projects/implicitron_trainer/configs/repro_singleseq_nerf_blender.yaml

@@ -8,9 +8,9 @@ data_source_ImplicitronDataSource_args:
   dataset_map_provider_class_type: BlenderDatasetMapProvider
   dataset_map_provider_BlenderDatasetMapProvider_args:
     base_dir: ${oc.env:BLENDER_DATASET_ROOT}
+    n_known_frames_for_test: null
     object_name: ${oc.env:BLENDER_SINGLESEQ_CLASS}
     path_manager_factory_class_type: PathManagerFactory
-    n_known_frames_for_test: null
     path_manager_factory_PathManagerFactory_args:
       silence_logs: true
 
@@ -20,8 +20,13 @@ model_factory_ImplicitronModelFactory_args:
       n_rays_per_image_sampled_from_mask: 4096
       scene_extent: 2.0
     renderer_MultiPassEmissionAbsorptionRenderer_args:
+      density_noise_std_train: 1.0
       n_pts_per_ray_fine_training: 128
       n_pts_per_ray_fine_evaluation: 128
+      raymarcher_EmissionAbsorptionRaymarcher_args:
+        blend_output: true
+        bg_color:
+        - 1.0
     loss_weights:
       loss_rgb_mse: 1.0
       loss_prev_stage_rgb_mse: 1.0
@@ -30,9 +35,12 @@ model_factory_ImplicitronModelFactory_args:
       loss_autodecoder_norm: 0.00
 
 optimizer_factory_ImplicitronOptimizerFactory_args:
-  exponential_lr_step_size: 2000
+  exponential_lr_step_size: 3001
+  lr_policy: Exponential
 
 training_loop_ImplicitronTrainingLoop_args:
-  max_epochs: 2000
+  max_epochs: 3001
-  visualize_interval: 0
+  metric_print_interval: 100
-  validation_interval: 30
+  store_checkpoints_purge: 3
+  test_when_finished: true
+  validation_interval: 100

projects/implicitron_trainer/tests/test_experiment.py

@@ -161,7 +161,7 @@ class TestExperiment(unittest.TestCase):
 
 
 class TestNerfRepro(unittest.TestCase):
-    @unittest.skip("This test reproduces full NERF training.")
+    @unittest.skip("This runs full NeRF training on Blender data.")
     def test_nerf_blender(self):
         # Train vanilla NERF.
         # Set env vars BLENDER_DATASET_ROOT and BLENDER_SINGLESEQ_CLASS first!

pytorch3d/implicitron/evaluation/evaluate_new_view_synthesis.py

@@ -242,14 +242,24 @@ def eval_batch(
     if frame_data.depth_map is None or frame_data.depth_map.sum() <= 0:
         warnings.warn("Empty or missing depth map in evaluation!")
 
+    if frame_data.mask_crop is None:
+        warnings.warn("mask_crop is None, assuming the whole image is valid.")
+
     # threshold the masks to make ground truth binary masks
-    mask_fg, mask_crop = [
+    # pyre-ignore [58]
-        (getattr(frame_data, k) >= mask_thr) for k in ("fg_probability", "mask_crop")
+    mask_fg = frame_data.fg_probability >= mask_thr
-    ]
+    mask_crop = (
+        frame_data.mask_crop
+        if frame_data.mask_crop is not None
+        # pyre-ignore [6]
+        else torch.ones_like(mask_fg)
+    )
+
     image_rgb_masked = mask_background(
         # pyre-fixme[6]: Expected `Tensor` for 1st param but got
         #  `Optional[torch.Tensor]`.
         frame_data.image_rgb,
+        # pyre-ignore [6]
         mask_fg,
         bg_color=bg_color,
     )
@@ -274,7 +284,7 @@ def eval_batch(
 
     results["iou"] = iou(
         cloned_render["mask_render"],
-        mask_fg,
+        mask_fg,  # pyre-ignore [6]
         mask=mask_crop,
     )
 
@@ -297,7 +307,7 @@ def eval_batch(
                     results[metric_name].item(), metric_name, loss_mask_now
                 )
 
-        if name_postfix == "_fg":
+        if name_postfix == "_fg" and frame_data.depth_map is not None:
             # only record depth metrics for the foreground
             _, abs_ = eval_depth(
                 cloned_render["depth_render"],
@@ -313,9 +323,7 @@ def eval_batch(
             if visualize:
                 visualizer.show_depth(abs_.mean().item(), name_postfix, loss_mask_now)
                 if break_after_visualising:
-                    import pdb  # noqa: B602
+                    breakpoint()  # noqa: B601
-
-                    pdb.set_trace()
 
     if lpips_model is not None:
         im1, im2 = [

pytorch3d/implicitron/evaluation/evaluator.py

@@ -126,6 +126,7 @@ class ImplicitronEvaluator(EvaluatorBase):
         )
 
         results = category_result["results"]
+        evaluate.pretty_print_nvs_metrics(results)
         if dump_to_json:
             _dump_to_json(epoch, exp_dir, results)
 
@@ -140,7 +141,6 @@ def _dump_to_json(
             r["eval_epoch"] = int(epoch)
     logger.info("Evaluation results")
 
-    evaluate.pretty_print_nvs_metrics(results)
     if exp_dir is None:
         raise ValueError("Cannot save results to json without a specified save path.")
     with open(os.path.join(exp_dir, "results_test.json"), "w") as f:
@@ -156,6 +156,7 @@ def _get_eval_frame_data(frame_data: Any) -> Any:
         frame_data.image_rgb
     )[:, None, None, None]
     for k in ("image_rgb", "depth_map", "fg_probability", "mask_crop"):
-        value_masked = getattr(frame_data_for_eval, k).clone() * is_known
+        value = getattr(frame_data_for_eval, k)
+        value_masked = value.clone() * is_known if value is not None else None
         setattr(frame_data_for_eval, k, value_masked)
     return frame_data_for_eval