Summary: One way to tidy the installation so we don't install files in site-packages/projects. Fixes https://github.com/facebookresearch/pytorch3d/issues/1279
Reviewed By: shapovalov, davnov134
Differential Revision: D38426772
fbshipit-source-id: ac1a54fbf230adb53904701e1f38bf9567f647ce
Summary: Don't copy from one part of config to another, rather do the copy within GenericModel.
Reviewed By: davnov134
Differential Revision: D38248828
fbshipit-source-id: ff8af985c37ea1f7df9e0aa0a45a58df34c3f893
Summary: Made the config system call open_dict when it calls the tweak function.
Reviewed By: shapovalov
Differential Revision: D38315334
fbshipit-source-id: 5924a92d8d0bf399bbf3788247f81fc990e265e7
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: Before this diff, train_stats.py would not be created by default, EXCEPT when resuming training. This makes them appear from start.
Reviewed By: shapovalov
Differential Revision: D38320341
fbshipit-source-id: 8ea5b99ec81c377ae129f58e78dc2eaff94821ad
Summary: Remove the dataset's need to provide the task type.
Reviewed By: davnov134, kjchalup
Differential Revision: D38314000
fbshipit-source-id: 3805d885b5d4528abdc78c0da03247edb9abf3f7
Summary:
Added _NEED_CONTROL
to JsonIndexDatasetMapProviderV2 and made dataset_tweak_args use it.
Reviewed By: bottler
Differential Revision: D38313914
fbshipit-source-id: 529847571065dfba995b609a66737bd91e002cfe
Summary: Only import it if you ask for it.
Reviewed By: kjchalup
Differential Revision: D38327167
fbshipit-source-id: 3f05231f26eda582a63afc71b669996342b0c6f9
Summary: Made eval_batches be set in call to `__init__` not after the construction as they were before
Reviewed By: bottler
Differential Revision: D38275943
fbshipit-source-id: 32737401d1ddd16c284e1851b7a91f8b041c406f
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: Make a dummy single-scene dataset using the code from generate_cow_renders (used in existing NeRF tutorials)
Reviewed By: kjchalup
Differential Revision: D38116910
fbshipit-source-id: 8db6df7098aa221c81d392e5cd21b0e67f65bd70
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:
This is an internal change in the config systen. It allows redefining a pluggable implementation with new default values. This is useful in notebooks / interactive use. For example, this now works.
class A(ReplaceableBase):
pass
registry.register
class B(A):
i: int = 4
class C(Configurable):
a: A
a_class_type: str = "B"
def __post_init__(self):
run_auto_creation(self)
expand_args_fields(C)
registry.register
class B(A):
i: int = 5
c = C()
assert c.a.i == 5
Reviewed By: shapovalov
Differential Revision: D38219371
fbshipit-source-id: 72911a9bd3426d3359cf8802cc016fc7f6d7713b
Summary:
Adding MeshRasterizerOpenGL, a faster alternative to MeshRasterizer. The new rasterizer follows the ideas from "Differentiable Surface Rendering via non-Differentiable Sampling".
The new rasterizer 20x faster on a 2M face mesh (try pose optimization on Nefertiti from https://www.cs.cmu.edu/~kmcrane/Projects/ModelRepository/!). The larger the mesh, the larger the speedup.
There are two main disadvantages:
* The new rasterizer works with an OpenGL backend, so requires pycuda.gl and pyopengl installed (though we avoided writing any C++ code, everything is in Python!)
* The new rasterizer is non-differentiable. However, you can still differentiate the rendering function if you use if with the new SplatterPhongShader which we recently added to PyTorch3D (see the original paper cited above).
Reviewed By: patricklabatut, jcjohnson
Differential Revision: D37698816
fbshipit-source-id: 54d120639d3cb001f096237807e54aced0acda25
Summary:
EGLContext is a utility to render with OpenGL without an attached display (that is, without a monitor).
DeviceContextManager allows us to avoid unnecessary context creations and releases. See docstrings for more info.
Reviewed By: jcjohnson
Differential Revision: D36562551
fbshipit-source-id: eb0d2a2f85555ee110e203d435a44ad243281d2c
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: We especially need omegaconf when testing impicitron.
Reviewed By: patricklabatut
Differential Revision: D37921440
fbshipit-source-id: 4e66fde35aa29f60eabd92bf459cd584cfd7e5ca
Summary:
X-link: https://github.com/fairinternal/pytorch3d/pull/39
Blender and LLFF cameras were sending screen space focal length and principal point to a camera init function expecting NDC
Reviewed By: shapovalov
Differential Revision: D37788686
fbshipit-source-id: 2ddf7436248bc0d174eceb04c288b93858138582
Summary: Add the conditioning types to the repro yaml files. In particular, this fixes test_conditioning_type.
Reviewed By: shapovalov
Differential Revision: D37914537
fbshipit-source-id: 621390f329d9da662d915eb3b7bc709206a20552
Summary:
I tried to run `experiment.py` and `pytorch3d_implicitron_runner` and faced the failure with this traceback: https://www.internalfb.com/phabricator/paste/view/P515734086
It seems to be due to the new release of OmegaConf (version=2.2.2) which requires different typing. This fix helped to overcome it.
Reviewed By: bottler
Differential Revision: D37881644
fbshipit-source-id: be0cd4ced0526f8382cea5bdca9b340e93a2fba2
Summary:
EPnP fails the test when the number of points is below 6. As suggested, quadratic option is in theory to deal with as few as 4 points (so num_pts_thresh=3 is set). And when num_pts > num_pts_thresh=4, skip_q is False.
To avoid bumping num_pts_thresh while passing all the original tests, check_output is set to False when num_pts < 6, similar to the logic in Line 123-127. It makes sure that the algo doesn't crash.
Reviewed By: shapovalov
Differential Revision: D37804438
fbshipit-source-id: 74576d63a9553e25e3ec344677edb6912b5f9354
Summary: Removing 1 from the crop mask does not seem sensible.
Reviewed By: bottler, shapovalov
Differential Revision: D37843680
fbshipit-source-id: 70cec80f9ea26deac63312da62b9c8af27d2a010
Summary:
1. Random sampling of num batches without replacement not supported.
2.Providers should implement the interface for the training loop to work.
Reviewed By: bottler, davnov134
Differential Revision: D37815388
fbshipit-source-id: 8a2795b524e733f07346ffdb20a9c0eb1a2b8190
Summary: Accelerate is an additional implicitron dependency, so document it.
Reviewed By: shapovalov
Differential Revision: D37786933
fbshipit-source-id: 11024fe604107881f8ca29e17cb5cbfe492fc7f9
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: one more bugfix in JsonIndexDataset
Reviewed By: bottler
Differential Revision: D37789138
fbshipit-source-id: 2fb2bda7448674091ff6b279175f0bbd16ff7a62
Summary:
This fixes a indexing bug in HardDepthShader and adds proper unit tests for both of the DepthShaders. This bug was introduced when updating the shader sizes and discovered when I switched my local model onto pytorch3d trunk instead of the patched copy.
Pull Request resolved: https://github.com/facebookresearch/pytorch3d/pull/1252
Test Plan:
Unit test + custom model code
```
pytest tests/test_shader.py
```
Reviewed By: bottler
Differential Revision: D37775767
Pulled By: d4l3k
fbshipit-source-id: 5f001903985976d7067d1fa0a3102d602790e3e8
Summary:
For 3D segmentation problems it's really useful to be able to train the models from multiple viewpoints using Pytorch3D as the renderer. Currently due to hardcoded assumptions in a few spots the mesh renderer only supports rendering RGB (3 dimensional) data. You can encode the classification information as 3 channel data but if you have more than 3 classes you're out of luck.
This relaxes the assumptions to make rendering semantic classes work with `HardFlatShader` and `AmbientLights` with no diffusion/specular. The other shaders/lights don't make any sense for classification since they mutate the texture values in some way.
This only requires changes in `Materials` and `AmbientLights`. The bulk of the code is the unit test.
Pull Request resolved: https://github.com/facebookresearch/pytorch3d/pull/1248
Test Plan: Added unit test that renders a 5 dimensional texture and compare dimensions 2-5 to a stored picture.
Reviewed By: bottler
Differential Revision: D37764610
Pulled By: d4l3k
fbshipit-source-id: 031895724d9318a6f6bab5b31055bb3f438176a5
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
