Mods and bugfixes for LLFF and Blender repros

Summary:
LLFF (and most/all non-synth datasets) will have no background/foreground distinction. Add support for data with no fg mask.

Also, we had a bug in stats loading, like this:
  * Load stats
  * One of the stats has a history of length 0
  * That's fine, e.g. maybe it's fg_error but the dataset has no notion of fg/bg. So leave it as len 0
  * Check whether all the stats have the same history length as an arbitrarily chosen "reference-stat"
  * Ooops the reference-stat happened to be the stat with length 0
  * assert (legit_stat_len == reference_stat_len (=0)) ---> failed assert

Also some minor fixes (from Jeremy's other diff) to support LLFF

Reviewed By: davnov134

Differential Revision: D38475272

fbshipit-source-id: 5b35ac86d1d5239759f537621f41a3aa4eb3bd68

projects/implicitron_trainer/configs/repro_singleseq_nerf_blender.yaml

@@ -1,7 +1,7 @@
 defaults:
 - repro_singleseq_base
 - _self_
-exp_dir: "./data/nerf_blender_publ/${oc.env:BLENDER_SINGLESEQ_CLASS}"
+exp_dir: "./data/nerf_blender_repro/${oc.env:BLENDER_SINGLESEQ_CLASS}"
 data_source_ImplicitronDataSource_args:
   data_loader_map_provider_SequenceDataLoaderMapProvider_args:
     dataset_length_train: 100
@@ -16,17 +16,18 @@ data_source_ImplicitronDataSource_args:
 
 model_factory_ImplicitronModelFactory_args:
   model_GenericModel_args:
-    raysampler_AdaptiveRaySampler_args:
+    mask_images: false
+    raysampler_class_type: NearFarRaySampler
+    raysampler_NearFarRaySampler_args:
       n_rays_per_image_sampled_from_mask: 4096
-      scene_extent: 2.0
+      min_depth: 2
+      max_depth: 6
     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
+        blend_output: false
     loss_weights:
       loss_rgb_mse: 1.0
       loss_prev_stage_rgb_mse: 1.0
@@ -35,11 +36,11 @@ model_factory_ImplicitronModelFactory_args:
       loss_autodecoder_norm: 0.00
 
 optimizer_factory_ImplicitronOptimizerFactory_args:
-  exponential_lr_step_size: 3001
+  exponential_lr_step_size: 2500
   lr_policy: Exponential
 
 training_loop_ImplicitronTrainingLoop_args:
-  max_epochs: 3001
+  max_epochs: 2000
   metric_print_interval: 100
   store_checkpoints_purge: 3
   test_when_finished: true

projects/implicitron_trainer/impl/training_loop.py

@@ -249,6 +249,7 @@ class ImplicitronTrainingLoop(TrainingLoopBase):  # pyre-ignore [13]
         stats = Stats(
             # log_vars should be a list, but OmegaConf might load them as ListConfig
             list(log_vars),
+            plot_file=os.path.join(exp_dir, "train_stats.pdf"),
             visdom_env=visdom_env_charts,
             verbose=False,
             visdom_server=self.visdom_server,

projects/implicitron_trainer/tests/experiment.yaml

@@ -95,6 +95,7 @@ data_source_ImplicitronDataSource_args:
     n_known_frames_for_test: null
     path_manager_factory_PathManagerFactory_args:
       silence_logs: true
+    downscale_factor: 4
   dataset_map_provider_RenderedMeshDatasetMapProvider_args:
     num_views: 40
     data_file: null

projects/implicitron_trainer/tests/test_experiment.py

@@ -162,7 +162,7 @@ class TestExperiment(unittest.TestCase):
 
 
 class TestNerfRepro(unittest.TestCase):
-    @unittest.skip("This runs full NeRF training on Blender data.")
+    @unittest.skip("This test runs full blender training.")
     def test_nerf_blender(self):
         # Train vanilla NERF.
         # Set env vars BLENDER_DATASET_ROOT and BLENDER_SINGLESEQ_CLASS first!
@@ -174,6 +174,22 @@ class TestNerfRepro(unittest.TestCase):
             experiment.dump_cfg(cfg)
             experiment_runner.run()
 
+    @unittest.skip("This test runs full llff training.")
+    def test_nerf_llff(self):
+        # Train vanilla NERF.
+        # Set env vars LLFF_DATASET_ROOT and LLFF_SINGLESEQ_CLASS first!
+        LLFF_SINGLESEQ_CLASS = os.environ["LLFF_SINGLESEQ_CLASS"]
+        if not interactive_testing_requested():
+            return
+        with initialize_config_dir(config_dir=str(IMPLICITRON_CONFIGS_DIR)):
+            cfg = compose(
+                config_name=f"repro_singleseq_nerf_llff_{LLFF_SINGLESEQ_CLASS}",
+                overrides=[],
+            )
+            experiment_runner = experiment.Experiment(**cfg)
+            experiment.dump_cfg(cfg)
+            experiment_runner.run()
+
     @unittest.skip("This test checks resuming of the NeRF training.")
     def test_nerf_blender_resume(self):
         # Train one train batch of NeRF, then resume for one more batch.