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
Summary: We don’t see much value in reporting metrics by camera difficulty while supporting that in new datasets is quite painful, hence deprecating training cameras in the data API and ignoring in evaluation.
Reviewed By: bottler
Differential Revision: D42678879
fbshipit-source-id: aad511f6cb2ca82745f31c19594e1d80594b61d7
Summary: Make Implicitron run without visdom installed.
Reviewed By: shapovalov
Differential Revision: D40587974
fbshipit-source-id: dc319596c7a4d10a4c54c556dabc89ad9d25c2fb
Summary: Loads the whole dataset and moves it to the device and sends it to for sampling to enable full dataset heterogeneous raysampling.
Reviewed By: bottler
Differential Revision: D39263009
fbshipit-source-id: c527537dfc5f50116849656c9e171e868f6845b1
Summary:
Changed ray_sampler and metrics to be able to use mixed frame raysampling.
Ray_sampler now has a new member which it passes to the pytorch3d raysampler.
If the raybundle is heterogeneous metrics now samples images by padding xys first. This reduces memory consumption.
Reviewed By: bottler, kjchalup
Differential Revision: D39542221
fbshipit-source-id: a6fec23838d3049ae5c2fd2e1f641c46c7c927e3
Summary: Workaround for oddity with new hydra.
Reviewed By: davnov134
Differential Revision: D39280639
fbshipit-source-id: 76e91947f633589945446db93cf2dbc259642f8a
Summary:
Adds yaml configs to train selected methods on CO3Dv2.
Few more updates:
1) moved some fields to base classes so that we can check is_multisequence in experiment.py
2) skip loading all train cameras for multisequence datasets, without this, co3d-fewview is untrainable
3) fix bug in json index dataset provider v2
Reviewed By: kjchalup
Differential Revision: D38952755
fbshipit-source-id: 3edac6fc8e20775aa70400bd73a0e6d52b091e0c
Summary:
Stats are logically connected to the training loop, not to the model. Hence, moving to the training loop.
Also removing resume_epoch from OptimizerFactory in favor of a single place - ModelFactory. This removes the need for config consistency checks etc.
Reviewed By: kjchalup
Differential Revision: D38313475
fbshipit-source-id: a1d188a63e28459df381ff98ad8acdcdb14887b7
Summary: Remove the dataset's need to provide the task type.
Reviewed By: davnov134, kjchalup
Differential Revision: D38314000
fbshipit-source-id: 3805d885b5d4528abdc78c0da03247edb9abf3f7
Summary: Currently, seeds are set only inside the train loop. But this does not ensure that the model weights are initialized the same way everywhere which makes all experiments irreproducible. This diff fixes it.
Reviewed By: bottler
Differential Revision: D38315840
fbshipit-source-id: 3d2ecebbc36072c2b68dd3cd8c5e30708e7dd808
Summary:
This large diff rewrites a significant portion of Implicitron's config hierarchy. The new hierarchy, and some of the default implementation classes, are as follows:
```
Experiment
data_source: ImplicitronDataSource
dataset_map_provider
data_loader_map_provider
model_factory: ImplicitronModelFactory
model: GenericModel
optimizer_factory: ImplicitronOptimizerFactory
training_loop: ImplicitronTrainingLoop
evaluator: ImplicitronEvaluator
```
1) Experiment (used to be ExperimentConfig) is now a top-level Configurable and contains as members mainly (mostly new) high-level factory Configurables.
2) Experiment's job is to run factories, do some accelerate setup and then pass the results to the main training loop.
3) ImplicitronOptimizerFactory and ImplicitronModelFactory are new high-level factories that create the optimizer, scheduler, model, and stats objects.
4) TrainingLoop is a new configurable that runs the main training loop and the inner train-validate step.
5) Evaluator is a new configurable that TrainingLoop uses to run validation/test steps.
6) GenericModel is not the only model choice anymore. Instead, ImplicitronModelBase (by default instantiated with GenericModel) is a member of Experiment and can be easily replaced by a custom implementation by the user.
All the new Configurables are children of ReplaceableBase, and can be easily replaced with custom implementations.
In addition, I added support for the exponential LR schedule, updated the config files and the test, as well as added a config file that reproduces NERF results and a test to run the repro experiment.
Reviewed By: bottler
Differential Revision: D37723227
fbshipit-source-id: b36bee880d6aa53efdd2abfaae4489d8ab1e8a27
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
Summary:
1. Respecting `visdom_show_preds` parameter when it is False.
2. Clipping the images pre-visualisation, which is important for methods like SRN that are not arare of pixel value range.
Reviewed By: bottler
Differential Revision: D37786439
fbshipit-source-id: 8dbb5104290bcc5c2829716b663cae17edc911bd
Summary:
## Changes:
- Added Accelerate Library and refactored experiment.py to use it
- Needed to move `init_optimizer` and `ExperimentConfig` to a separate file to be compatible with submitit/hydra
- Needed to make some modifications to data loaders etc to work well with the accelerate ddp wrappers
- Loading/saving checkpoints incorporates an unwrapping step so remove the ddp wrapped model
## Tests
Tested with both `torchrun` and `submitit/hydra` on two gpus locally. Here are the commands:
**Torchrun**
Modules loaded:
```sh
1) anaconda3/2021.05 2) cuda/11.3 3) NCCL/2.9.8-3-cuda.11.3 4) gcc/5.2.0. (but unload gcc when using submit)
```
```sh
torchrun --nnodes=1 --nproc_per_node=2 experiment.py --config-path ./configs --config-name repro_singleseq_nerf_test
```
**Submitit/Hydra Local test**
```sh
~/pytorch3d/projects/implicitron_trainer$ HYDRA_FULL_ERROR=1 python3.9 experiment.py --config-name repro_singleseq_nerf_test --multirun --config-path ./configs hydra/launcher=submitit_local hydra.launcher.gpus_per_node=2 hydra.launcher.tasks_per_node=2 hydra.launcher.nodes=1
```
**Submitit/Hydra distributed test**
```sh
~/implicitron/pytorch3d$ python3.9 experiment.py --config-name repro_singleseq_nerf_test --multirun --config-path ./configs hydra/launcher=submitit_slurm hydra.launcher.gpus_per_node=8 hydra.launcher.tasks_per_node=8 hydra.launcher.nodes=1 hydra.launcher.partition=learnlab hydra.launcher.timeout_min=4320
```
## TODOS:
- Fix distributed evaluation: currently this doesn't work as the input format to the evaluation function is not suitable for gathering across gpus (needs to be nested list/tuple/dicts of objects that satisfy `is_torch_tensor`) and currently `frame_data` contains `Cameras` type.
- Refactor the `accelerator` object to be accessible by all functions instead of needing to pass it around everywhere? Maybe have a `Trainer` class and add it as a method?
- Update readme with installation instructions for accelerate and also commands for running jobs with torchrun and submitit/hydra
X-link: https://github.com/fairinternal/pytorch3d/pull/37
Reviewed By: davnov134, kjchalup
Differential Revision: D37543870
Pulled By: bottler
fbshipit-source-id: be9eb4e91244d4fe3740d87dafec622ae1e0cf76
Summary: As part of removing Task, move camera difficulty bin breaks from hard code to the top level.
Reviewed By: davnov134
Differential Revision: D37491040
fbshipit-source-id: f2d6775ebc490f6f75020d13f37f6b588cc07a0b
Summary: As part of removing Task, make the dataset code generate the source cameras for itself. There's a small optimization available here, in that the JsonIndexDataset could avoid loading images.
Reviewed By: shapovalov
Differential Revision: D37313423
fbshipit-source-id: 3e5e0b2aabbf9cc51f10547a3523e98c72ad8755
Summary: Copy code from NeRF for loading LLFF data and blender synthetic data, and create dataset objects for them
Reviewed By: shapovalov
Differential Revision: D35581039
fbshipit-source-id: af7a6f3e9a42499700693381b5b147c991f57e5d
Summary: The ImplicitronDataset class corresponds to JsonIndexDatasetMapProvider
Reviewed By: shapovalov
Differential Revision: D36661396
fbshipit-source-id: 80ca2ff81ef9ecc2e3d1f4e1cd14b6f66a7ec34d
Summary: replace dataset_zoo with a pluggable DatasetMapProvider. The logic is now in annotated_file_dataset_map_provider.
Reviewed By: shapovalov
Differential Revision: D36443965
fbshipit-source-id: 9087649802810055e150b2fbfcc3c197a761f28a
Summary: Separate ImplicitronDatasetBase and FrameData (to be used by all data sources) from ImplicitronDataset (which is specific).
Reviewed By: shapovalov
Differential Revision: D36413111
fbshipit-source-id: 3725744cde2e08baa11aff4048237ba10c7efbc6
Summary:
Move dataset_args and dataloader_args from ExperimentConfig into a new member called datasource so that it can contain replaceables.
Also add enum Task for task type.
Reviewed By: shapovalov
Differential Revision: D36201719
fbshipit-source-id: 47d6967bfea3b7b146b6bbd1572e0457c9365871
Summary:
To avoid model_zoo, we need to make GenericModel pluggable.
I also align creation APIs for convenience.
Reviewed By: bottler, davnov134
Differential Revision: D35933093
fbshipit-source-id: 8228926528eb41a795fbfbe32304b8019197e2b1
Summary:
Try again to solve https://github.com/facebookresearch/pytorch3d/issues/1144 pickling problem.
D35258561 (24260130ce) didn't work.
When writing a function or vanilla class C which you want people to be able to call get_default_args on, you must add the line enable_get_default_args(C) to it. This causes autogeneration of a hidden dataclass in the module.
Reviewed By: davnov134
Differential Revision: D35364410
fbshipit-source-id: 53f6e6fff43e7142ae18ca3b06de7d0c849ef965
Summary:
ListConfig and DictConfig members of get_default_args(X) when X is a callable will contain references to a temporary dataclass and therefore be unpicklable. Avoid this in a few cases.
Fixes https://github.com/facebookresearch/pytorch3d/issues/1144
Reviewed By: shapovalov
Differential Revision: D35258561
fbshipit-source-id: e52186825f52accee9a899e466967a4ff71b3d25
Summary:
Before the fix, running get_default_args(C: Callable) returns an unstructured DictConfig which causes Enums to be handled incorrectly. This is a fix.
WIP update: Currently tests still fail whenever a function signature contains an untyped argument: This needs to be somehow fixed.
Reviewed By: bottler
Differential Revision: D34932124
fbshipit-source-id: ecdc45c738633cfea5caa7480ba4f790ece931e8
