Add the OverfitModel

Summary:
Introduces the OverfitModel for NeRF-style training with overfitting to one scene.
It is a specific case of GenericModel. It has been disentangle to ease usage.

## General modification

1. Modularize a minimum GenericModel to introduce OverfitModel
2. Introduce OverfitModel and ensure through unit testing that it behaves like GenericModel.

## Modularization

The following methods have been extracted from GenericModel to allow modularity with ManyViewModel:
- get_objective is now a call to weighted_sum_losses
- log_loss_weights
- prepare_inputs

The generic methods have been moved to an utils.py file.

Simplify the code to introduce OverfitModel.

Private methods like chunk_generator are now public and can now be used by ManyViewModel.

Reviewed By: shapovalov

Differential Revision: D43771992

fbshipit-source-id: 6102aeb21c7fdd56aa2ff9cd1dd23fd9fbf26315

Author: EmGarr

projects/implicitron_trainer/README.md

@@ -248,7 +248,7 @@ The main object for this trainer loop is `Experiment`. It has four top-level rep
 * `data_source`: This is a `DataSourceBase` which defaults to `ImplicitronDataSource`.
 It constructs the data sets and dataloaders.
 * `model_factory`: This is a `ModelFactoryBase` which defaults to `ImplicitronModelFactory`.
-It constructs the model, which is usually an instance of implicitron's main `GenericModel` class, and can load its weights from a checkpoint.
+It constructs the model, which is usually an instance of `OverfitModel` (for NeRF-style training with overfitting to one scene) or `GenericModel` (that is able to generalize to multiple scenes by NeRFormer-style conditioning on other scene views), and can load its weights from a checkpoint.
 * `optimizer_factory`: This is an `OptimizerFactoryBase` which defaults to `ImplicitronOptimizerFactory`.
 It constructs the optimizer and can load its weights from a checkpoint.
 * `training_loop`: This is a `TrainingLoopBase` which defaults to `ImplicitronTrainingLoop` and defines the main training loop.
@@ -292,6 +292,43 @@ model_GenericModel_args: GenericModel
         ╘== ReductionFeatureAggregator
 ```
 
+Here is the class structure of OverfitModel:
+
+```
+model_OverfitModel_args: OverfitModel
+└-- raysampler_*_args: RaySampler
+    ╘== AdaptiveRaysampler
+    ╘== NearFarRaysampler
+└-- renderer_*_args: BaseRenderer
+    ╘== MultiPassEmissionAbsorptionRenderer
+    ╘== LSTMRenderer
+    ╘== SignedDistanceFunctionRenderer
+        └-- ray_tracer_args: RayTracing
+        └-- ray_normal_coloring_network_args: RayNormalColoringNetwork
+└-- implicit_function_*_args: ImplicitFunctionBase
+    ╘== NeuralRadianceFieldImplicitFunction
+    ╘== SRNImplicitFunction
+        └-- raymarch_function_args: SRNRaymarchFunction
+        └-- pixel_generator_args: SRNPixelGenerator
+    ╘== SRNHyperNetImplicitFunction
+        └-- hypernet_args: SRNRaymarchHyperNet
+        └-- pixel_generator_args: SRNPixelGenerator
+    ╘== IdrFeatureField
+└-- coarse_implicit_function_*_args: ImplicitFunctionBase
+    ╘== NeuralRadianceFieldImplicitFunction
+    ╘== SRNImplicitFunction
+        └-- raymarch_function_args: SRNRaymarchFunction
+        └-- pixel_generator_args: SRNPixelGenerator
+    ╘== SRNHyperNetImplicitFunction
+        └-- hypernet_args: SRNRaymarchHyperNet
+        └-- pixel_generator_args: SRNPixelGenerator
+    ╘== IdrFeatureField
+```
+
+OverfitModel has been introduced to create a simple class to disantagle Nerfs which the overfit pattern
+from the GenericModel.
+
+
 Please look at the annotations of the respective classes or functions for the lists of hyperparameters.
 `tests/experiment.yaml` shows every possible option if you have no user-defined classes.
 

projects/implicitron_trainer/configs/overfit_base.yaml

@@ -0,0 +1,79 @@
+defaults:
+- default_config
+- _self_
+exp_dir: ./data/exps/overfit_base/
+training_loop_ImplicitronTrainingLoop_args:
+  visdom_port: 8097
+  visualize_interval: 0
+  max_epochs: 1000
+data_source_ImplicitronDataSource_args:
+  data_loader_map_provider_class_type: SequenceDataLoaderMapProvider
+  dataset_map_provider_class_type: JsonIndexDatasetMapProvider
+  data_loader_map_provider_SequenceDataLoaderMapProvider_args:
+    dataset_length_train: 1000
+    dataset_length_val: 1
+    num_workers: 8
+  dataset_map_provider_JsonIndexDatasetMapProvider_args:
+    dataset_root: ${oc.env:CO3D_DATASET_ROOT}
+    n_frames_per_sequence: -1
+    test_on_train: true
+    test_restrict_sequence_id: 0
+    dataset_JsonIndexDataset_args:
+      load_point_clouds: false
+      mask_depths: false
+      mask_images: false
+model_factory_ImplicitronModelFactory_args:
+  model_class_type: "OverfitModel"
+  model_OverfitModel_args:
+    loss_weights:
+      loss_mask_bce: 1.0
+      loss_prev_stage_mask_bce: 1.0
+      loss_autodecoder_norm: 0.01
+      loss_rgb_mse: 1.0
+      loss_prev_stage_rgb_mse: 1.0
+    output_rasterized_mc: false
+    chunk_size_grid: 102400
+    render_image_height: 400
+    render_image_width: 400
+    share_implicit_function_across_passes: false
+    implicit_function_class_type: "NeuralRadianceFieldImplicitFunction"
+    implicit_function_NeuralRadianceFieldImplicitFunction_args:
+      n_harmonic_functions_xyz: 10
+      n_harmonic_functions_dir: 4
+      n_hidden_neurons_xyz: 256
+      n_hidden_neurons_dir: 128
+      n_layers_xyz: 8
+      append_xyz:
+      - 5
+    coarse_implicit_function_class_type: "NeuralRadianceFieldImplicitFunction"
+    coarse_implicit_function_NeuralRadianceFieldImplicitFunction_args:
+      n_harmonic_functions_xyz: 10
+      n_harmonic_functions_dir: 4
+      n_hidden_neurons_xyz: 256
+      n_hidden_neurons_dir: 128
+      n_layers_xyz: 8
+      append_xyz:
+      - 5
+    raysampler_AdaptiveRaySampler_args:
+      n_rays_per_image_sampled_from_mask: 1024
+      scene_extent: 8.0
+      n_pts_per_ray_training: 64
+      n_pts_per_ray_evaluation: 64
+      stratified_point_sampling_training: true
+      stratified_point_sampling_evaluation: false
+    renderer_MultiPassEmissionAbsorptionRenderer_args:
+      n_pts_per_ray_fine_training: 64
+      n_pts_per_ray_fine_evaluation: 64
+      append_coarse_samples_to_fine: true
+      density_noise_std_train: 1.0
+optimizer_factory_ImplicitronOptimizerFactory_args:
+  breed: Adam
+  weight_decay: 0.0
+  lr_policy: MultiStepLR
+  multistep_lr_milestones: []
+  lr: 0.0005
+  gamma: 0.1
+  momentum: 0.9
+  betas:
+  - 0.9
+  - 0.999

projects/implicitron_trainer/configs/overfit_singleseq_base.yaml

@@ -0,0 +1,42 @@
+defaults:
+- overfit_base
+- _self_
+data_source_ImplicitronDataSource_args:
+  data_loader_map_provider_SequenceDataLoaderMapProvider_args:
+    batch_size: 1
+    dataset_length_train: 1000
+    dataset_length_val: 1
+    num_workers: 8
+  dataset_map_provider_JsonIndexDatasetMapProvider_args:
+    assert_single_seq: true
+    n_frames_per_sequence: -1
+    test_restrict_sequence_id: 0
+    test_on_train: false
+model_factory_ImplicitronModelFactory_args:
+  model_class_type: "OverfitModel"
+  model_OverfitModel_args:
+    render_image_height: 800
+    render_image_width: 800
+    log_vars:
+    - loss_rgb_psnr_fg
+    - loss_rgb_psnr
+    - loss_eikonal
+    - loss_prev_stage_rgb_psnr
+    - loss_mask_bce
+    - loss_prev_stage_mask_bce
+    - loss_rgb_mse
+    - loss_prev_stage_rgb_mse
+    - loss_depth_abs
+    - loss_depth_abs_fg
+    - loss_kl
+    - loss_mask_neg_iou
+    - objective
+    - epoch
+    - sec/it
+optimizer_factory_ImplicitronOptimizerFactory_args:
+  lr: 0.0005
+  multistep_lr_milestones:
+  - 200
+  - 300
+training_loop_ImplicitronTrainingLoop_args:
+  max_epochs: 400

projects/implicitron_trainer/configs/overfit_singleseq_nerf_blender.yaml

@@ -0,0 +1,56 @@
+defaults:
+- overfit_singleseq_base
+- _self_
+exp_dir: "./data/overfit_nerf_blender_repro/${oc.env:BLENDER_SINGLESEQ_CLASS}"
+data_source_ImplicitronDataSource_args:
+  data_loader_map_provider_SequenceDataLoaderMapProvider_args:
+    dataset_length_train: 100
+  dataset_map_provider_class_type: BlenderDatasetMapProvider
+  dataset_map_provider_BlenderDatasetMapProvider_args:
+    base_dir: ${oc.env:BLENDER_DATASET_ROOT}/${oc.env:BLENDER_SINGLESEQ_CLASS}
+    n_known_frames_for_test: null
+    object_name: ${oc.env:BLENDER_SINGLESEQ_CLASS}
+    path_manager_factory_class_type: PathManagerFactory
+    path_manager_factory_PathManagerFactory_args:
+      silence_logs: true
+
+model_factory_ImplicitronModelFactory_args:
+  model_class_type: "OverfitModel"
+  model_OverfitModel_args:
+    mask_images: false
+    raysampler_class_type: AdaptiveRaySampler
+    raysampler_AdaptiveRaySampler_args:
+      n_pts_per_ray_training: 64
+      n_pts_per_ray_evaluation: 64
+      n_rays_per_image_sampled_from_mask: 4096
+      stratified_point_sampling_training: true
+      stratified_point_sampling_evaluation: false
+      scene_extent: 2.0
+      scene_center:
+      - 0.0
+      - 0.0
+      - 0.0
+    renderer_MultiPassEmissionAbsorptionRenderer_args:
+      density_noise_std_train: 0.0
+      n_pts_per_ray_fine_training: 128
+      n_pts_per_ray_fine_evaluation: 128
+      raymarcher_EmissionAbsorptionRaymarcher_args:
+        blend_output: false
+    loss_weights:
+      loss_rgb_mse: 1.0
+      loss_prev_stage_rgb_mse: 1.0
+      loss_mask_bce: 0.0
+      loss_prev_stage_mask_bce: 0.0
+      loss_autodecoder_norm: 0.00
+
+optimizer_factory_ImplicitronOptimizerFactory_args:
+  exponential_lr_step_size: 3001
+  lr_policy: LinearExponential
+  linear_exponential_lr_milestone: 200
+
+training_loop_ImplicitronTrainingLoop_args:
+  max_epochs: 6000
+  metric_print_interval: 10
+  store_checkpoints_purge: 3
+  test_when_finished: true
+  validation_interval: 100

projects/implicitron_trainer/experiment.py

@@ -59,7 +59,7 @@
     DataSourceBase,
     ImplicitronDataSource,
 )
-from pytorch3d.implicitron.models.generic_model import ImplicitronModelBase
+from pytorch3d.implicitron.models.base_model import ImplicitronModelBase
 
 from pytorch3d.implicitron.models.renderer.multipass_ea import (
     MultiPassEmissionAbsorptionRenderer,