Summary:
Adds the ability to have different learning rates for different parts of the model. The trainable parts of the implicitron have a new member
param_groups: dictionary where keys are names of individual parameters,
or module’s members and values are the parameter group where the
parameter/member will be sorted to. "self" key is used to denote the
parameter group at the module level. Possible keys, including the "self" key
do not have to be defined. By default all parameters are put into "default"
parameter group and have the learning rate defined in the optimizer,
it can be overriden at the:
- module level with “self” key, all the parameters and child
module s parameters will be put to that parameter group
- member level, which is the same as if the `param_groups` in that
member has key=“self” and value equal to that parameter group.
This is useful if members do not have `param_groups`, for
example torch.nn.Linear.
- parameter level, parameter with the same name as the key
will be put to that parameter group.
And in the optimizer factory, parameters and their learning rates are recursively gathered.
Reviewed By: shapovalov
Differential Revision: D40145802
fbshipit-source-id: 631c02b8d79ee1c0eb4c31e6e42dbd3d2882078a
Summary:
Small config system fix. Allows get_default_args to work on an instance which has been created with a dict (instead of a DictConfig) as an args field. E.g.
```
gm = GenericModel(
raysampler_AdaptiveRaySampler_args={"scene_extent": 4.0}
)
OmegaConf.structured(gm1)
```
Reviewed By: shapovalov
Differential Revision: D40341047
fbshipit-source-id: 587d0e8262e271df442a80858949a48e5d6db3df
Summary: Tensorf does relu or softmax after the density grid. This diff adds the ability to replicate that.
Reviewed By: bottler
Differential Revision: D40023228
fbshipit-source-id: 9f19868cd68460af98ab6e61c7f708158c26dc08
Summary:
TensoRF at step 2000 does volume croping and resizing.
At those steps it calculates part of the voxel grid which has density big enough to have objects and resizes the grid to fit that object.
Change is done on 3 levels:
- implicit function subscribes to epochs and at specific epochs finds the bounding box of the object and calls resizing of the color and density voxel grids to fit it
- VoxelGrid module calls cropping of the underlaying voxel grid and resizing to fit previous size it also adjusts its extends and translation to match wanted size
- Each voxel grid has its own way of cropping the underlaying data
Reviewed By: kjchalup
Differential Revision: D39854548
fbshipit-source-id: 5435b6e599aef1eaab980f5421d3369ee4829c50
Summary: Forward method is sped up using the scaffold, a low resolution voxel grid which is used to filter out the points in empty space. These points will be predicted as having 0 density and (0, 0, 0) color. The points which were not evaluated as empty space will be passed through the steps outlined above.
Reviewed By: kjchalup
Differential Revision: D39579671
fbshipit-source-id: 8eab8bb43ef77c2a73557efdb725e99a6c60d415
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: new implicitronRayBundle with added cameraIDs and camera counts. Added to enable a single raybundle inside Implicitron and easier extension in the future. Since RayBundle is named tuple and RayBundleHeterogeneous is dataclass and RayBundleHeterogeneous cannot inherit RayBundle. So if there was no ImplicitronRayBundle every function that uses RayBundle now would have to use Union[RayBundle, RaybundleHeterogeneous] which is confusing and unecessary complicated.
Reviewed By: bottler, kjchalup
Differential Revision: D39262999
fbshipit-source-id: ece160e32f6c88c3977e408e966789bf8307af59
Summary:
`get_rotation_to_best_fit_xy` is useful to expose externally, however there was a bug (which we probably did not care about for our use case): it could return a rotation matrix with det(R) == −1.
The diff fixes that, and also makes centroid optional (it can be computed from points).
Reviewed By: bottler
Differential Revision: D39926791
fbshipit-source-id: 5120c7892815b829f3ddcc23e93d4a5ec0ca0013
Summary: Any module can be subscribed to step updates from the training loop. Once the training loop publishes a step the voxel grid changes its dimensions. During the construction of VoxelGridModule and its parameters it does not know which is the resolution that will be loaded from checkpoint, so before the checkpoint loading a hook runs which changes the VoxelGridModule's parameters to match shapes of the loaded checkpoint.
Reviewed By: bottler
Differential Revision: D39026775
fbshipit-source-id: 0d359ea5c8d2eda11d773d79c7513c83585d5f17
Summary: Add the ability to process arbitrary point shapes `[n_grids, ..., 3]` instead of only `[n_grids, n_points, 3]`.
Reviewed By: bottler
Differential Revision: D39574373
fbshipit-source-id: 0a9ecafe9ea58cd8f909644de43a1185ecf934f4
Summary:
- indicate location of OmegaConf.structured failures
- split the data gathering from enable_get_default_args to ease experimenting with it.
- comment fixes.
- nicer error when a_class_type has weird type.
Reviewed By: kjchalup
Differential Revision: D39434447
fbshipit-source-id: b80c7941547ca450e848038ef5be95b7ebbe8f3e
Summary:
Move the flyaround rendering function into core implicitron.
The unblocks an example in the facebookresearch/co3d repo.
Reviewed By: bottler
Differential Revision: D39257801
fbshipit-source-id: 6841a88a43d4aa364dd86ba83ca2d4c3cf0435a4
Summary: Currently some implicit functions in implicitron take a raybundle, others take ray_points_world. raybundle is what they really need. However, the raybundle is going to become a bit more flexible later, as it will contain different numbers of rays for each camera.
Reviewed By: bottler
Differential Revision: D39173751
fbshipit-source-id: ebc038e426d22e831e67a18ba64655d8a61e1eb9
Summary: Added replacable decoding functions which will be applied after the voxel grid to get color and density
Reviewed By: bottler
Differential Revision: D38829763
fbshipit-source-id: f21ce206c1c19548206ea2ce97d7ebea3de30a23
Summary: Simple wrapper around voxel grids to make them a module
Reviewed By: bottler
Differential Revision: D38829762
fbshipit-source-id: dfee85088fa3c65e396cc7d3bf7ebaaffaadb646
Summary: Added voxel grid classes from TensoRF, both in their factorized (CP and VM) and full form.
Reviewed By: bottler
Differential Revision: D38465419
fbshipit-source-id: 8b306338af58dc50ef47a682616022a0512c0047
Summary:
generic_model_args no longer exists. Update some references to it, mostly in doc.
This fixes the testing of all the yaml files in test_forward pass.
Reviewed By: shapovalov
Differential Revision: D38789202
fbshipit-source-id: f11417efe772d7f86368b3598aa66c52b1309dbf
Summary: Previously, "psnr" was evaluated between the masked g.t. image and the render. To avoid confusion, "psnr" is now renamed to "psnr_masked".
Reviewed By: bottler
Differential Revision: D38707511
fbshipit-source-id: 8ee881ab1a05453d6692dde9782333a47d8c1234
Summary: Reports also the PSNR between the unmasked G.T. image and the render.
Reviewed By: bottler
Differential Revision: D38655943
fbshipit-source-id: 1603a2d02116ea1ce037e5530abe1afc65a2ba93
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
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:
Added _NEED_CONTROL
to JsonIndexDatasetMapProviderV2 and made dataset_tweak_args use it.
Reviewed By: bottler
Differential Revision: D38313914
fbshipit-source-id: 529847571065dfba995b609a66737bd91e002cfe
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 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:
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: A new version of json index dataset provider supporting CO3Dv2
Reviewed By: shapovalov
Differential Revision: D37690918
fbshipit-source-id: bf2d5fc9d0f1220259e08661dafc69cdbe6b7f94
Summary:
Implements several changes needed for the CO3Dv2 release:
- FrameData contains crop_bbox_xywh which defines the outline of the image crop corresponding to the image-shaped tensors in FrameData
- revised the definition of a bounding box inside JsonDatasetIndex: bbox_xyxy is [xmin, ymin, xmax, ymax], where xmax, ymax are not inclusive; bbox_xywh = [xmin, ymain, xmax-xmin, ymax-ymin]
- is_filtered for detecting whether the entries of the dataset were somehow filtered
- seq_frame_index_to_dataset_index allows to skip entries that are not present in the dataset
Reviewed By: shapovalov
Differential Revision: D37687547
fbshipit-source-id: 7842756b0517878cc0964fc0935d3c0769454d78
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:
Make ViewMetrics easy to replace by putting them into an OmegaConf dataclass.
Also, re-word a few variable names and fix minor TODOs.
Reviewed By: bottler
Differential Revision: D37327157
fbshipit-source-id: 78d8e39bbb3548b952f10abbe05688409fb987cc
Summary:
David had his code crashed when using frame_annot["meta"] dictionary. Turns out we had a typo.
The tests were passing by chance since all the keys were single-character strings.
Reviewed By: bottler
Differential Revision: D37503987
fbshipit-source-id: c12b0df21116cfbbc4675a0182b9b9e6d62bad2e
Summary: small followup to D37172537 (cba26506b6) and D37209012 (81d63c6382): changing default #harmonics and improving a test
Reviewed By: shapovalov
Differential Revision: D37412357
fbshipit-source-id: 1af1005a129425fd24fa6dd213d69c71632099a0