Summary:
Use aten instead of torch interface in all cuda code. This allows the cuda build to work with pytorch 1.5 with GCC 5 (e.g. the compiler of ubuntu 16.04LTS). This wasn't working. It has been failing with errors like the below, perhaps due to a bug in nvcc.
```
torch/include/torch/csrc/api/include/torch/nn/cloneable.h:68:61: error: invalid static_cast from type ‘const torch::OrderedDict<std::basic_string<char>, std::shared_ptr<torch::nn::Module> >’ to type ‘torch::OrderedDict<std::basic_string<char>, std::shared_ptr<torch::nn::Module> >
```
Reviewed By: nikhilaravi
Differential Revision: D21204029
fbshipit-source-id: ca6bdbcecf42493365e1c23a33fe35e1759fe8b6
Summary:
davnov134 found that the algorithm crashes if X is an axis-aligned plane. This is because I implemented scaling control points by `X.std()` as a poor man’s version of PCA whitening.
I checked that it does not bring consistent improvements, so let’s get rid of it.
The algorithm still results in slightly higher errors on the axis aligned planes but at least it does not crash. As a next step, I will experiment with detecting a planar case and using 3-point barycentric coordinates rather than 4-points.
Reviewed By: davnov134
Differential Revision: D21179968
fbshipit-source-id: 1f002fce5541934486b51808be0e910324977222
Summary:
We have multiple KNN CUDA implementations. From python, users can currently request a particular implementation via the `version` flag, but they have no way of knowing which implementations can be used for a given problem.
This diff exposes a function `pytorch3d._C.knn_check_version(version, D, K)` that returns whether a particular version can be used.
Reviewed By: nikhilaravi
Differential Revision: D21162573
fbshipit-source-id: 6061960bdcecba454fd920b00036f4e9ff3fdbc0
Summary:
Modify test_chamfer for more robustness. Avoid empty pointclouds, including where point_reduction is mean, for which we currently return nan (*), and so that we aren't looking at an empty gradient. Make sure we aren't using padding as points in the homogenous cases in the tests, which will lead to a tie between closest points and therefore a potential instability in the gradient - see https://github.com/pytorch/pytorch/issues/35699.
(*) This doesn't attempt to fix the nan.
Reviewed By: nikhilaravi, gkioxari
Differential Revision: D21157322
fbshipit-source-id: a609e84e25a24379c8928ff645d587552526e4af
Summary: cuda 10.2 location on linux. Also remove unused conda test dependencies.
Reviewed By: nikhilaravi
Differential Revision: D21176409
fbshipit-source-id: dd3f339a92233ff16877ba76506ddf8f4418715d
Summary:
Added backface culling as an option to the `raster_settings`. This is needed for the full forward rendering of shapenet meshes with texture (some meshes contain
multiple overlapping segments which have different textures).
For a triangle (v0, v1, v2) define the vectors A = (v1 - v0) and B = (v2 − v0) and use this to calculate the area of the triangle as:
```
area = 0.5 * A x B
area = 0.5 * ((x1 − x0)(y2 − y0) − (x2 − x0)(y1 − y0))
```
The area will be positive if (v0, v1, v2) are oriented counterclockwise (a front face), and negative if (v0, v1, v2) are oriented clockwise (a back face).
We can reuse the `edge_function` as it already calculates the triangle area.
Reviewed By: jcjohnson
Differential Revision: D20960115
fbshipit-source-id: 2d8a4b9ccfb653df18e79aed8d05c7ec0f057ab1
Summary:
Add conda packages for pytorch 1.5. Make wheels be only pytorch 1.5.
Note that pytorch 1.4 has conda packages for cuda 9.2, 10.0 and 10.1, whilst pytorch 1.5 has packages for cuda 9.2, 10.1 and 10.2. We mirror these choices.
Reviewed By: nikhilaravi
Differential Revision: D21157392
fbshipit-source-id: 2f7311e6a83774a6d6c8afb8110b8bd9f37f1454
Summary:
Fix a bug which resulted in a rendering artifacts if the image size was not a multiple of 16.
Fix: Revert coarse rasterization to original implementation and only update fine rasterization to reverse the ordering of Y and X axis. This is much simpler than the previous approach!
Additional changes:
- updated mesh rendering end-end tests to check outputs from both naive and coarse to fine rasterization.
- added pointcloud rendering end-end tests
Reviewed By: gkioxari
Differential Revision: D21102725
fbshipit-source-id: 2e7e1b013dd6dd12b3a00b79eb8167deddb2e89a
Summary:
None of the current test_build tests make sense during `conda build`.
Also remove the unnecessary dependency on the `six` library.
Reviewed By: nikhilaravi
Differential Revision: D20893852
fbshipit-source-id: 685f0446eaa0bd9151eeee89fc630a1ddc0252ff
Summary: This is mostly replacing the old PackedTensorAccessor with the new PackedTensorAccessor64.
Reviewed By: gkioxari
Differential Revision: D21088773
fbshipit-source-id: 5973e5a29d934eafb7c70ec5ec154ca076b64d27
Summary: A couple of files for the removed nearest_neighbor functionality are left behind.
Reviewed By: nikhilaravi
Differential Revision: D21088624
fbshipit-source-id: 4bb29016b4e5f63102765b384c363733b60032fa
Summary:
Efficient PnP algorithm to fit 2D to 3D correspondences under perspective assumption.
Benchmarked both variants of nullspace and pick one; SVD takes 7 times longer in the 100K points case.
Reviewed By: davnov134, gkioxari
Differential Revision: D20095754
fbshipit-source-id: 2b4519729630e6373820880272f674829eaed073
Summary: Made a CameraBase class. Added `unproject_points` method for each camera class.
Reviewed By: nikhilaravi
Differential Revision: D20373602
fbshipit-source-id: 7e3da5ae420091b5fcab400a9884ef29ad7a7343
Summary: Estimates normals of a point cloud.
Reviewed By: gkioxari
Differential Revision: D20860182
fbshipit-source-id: 652ec2743fa645e02c01ffa37c2971bf27b89cef
Summary: It seemed that even though the chamfer diff was rebased on top of the knn autograd diff, some of the final updates did not get applied. I'm really surprised that the sandcastle tests did not fail and prevent the diff from landing.
Reviewed By: gkioxari
Differential Revision: D21066156
fbshipit-source-id: 5216efe95180c1b6082d0bac404fa1920cfb7b02
Summary: knn is more general and faster than the nearest_neighbor code, so remove the latter.
Reviewed By: gkioxari
Differential Revision: D20816424
fbshipit-source-id: 75d6c44d17180752d0c9859814bbdf7892558158
Summary:
Allow Pointclouds objects and heterogenous data to be provided for Chamfer loss. Remove "none" as an option for point_reduction because it doesn't make sense and in the current implementation is effectively the same as "sum".
Possible improvement: create specialised operations for sum and cosine_similarity of padded tensors, to avoid having to create masks. sum would be useful elsewhere.
Reviewed By: gkioxari
Differential Revision: D20816301
fbshipit-source-id: 0f32073210225d157c029d80de450eecdb64f4d2
Summary: Remove `bin_size` and `max_faces_per_pixel` from being specified. This means the coarse-to-fine rasterization will be used by default and will help avoid confusion with the naive version.
Reviewed By: jcjohnson
Differential Revision: D20908905
fbshipit-source-id: c181c88e844d888aa81a36870918307961dc1175
Summary:
Pytorch 1.5 is coming soon. I imagine we will want the ability to upload conda packages for pytorch3d to anaconda cloud for each of pytorch 1.4 and pytorch 1.5. This change adds the dependent pytorch version to the name of the conda package to make that feasible.
As an example, a built package after this change will have a name like `linux-64/pytorch3d-0.1.1-py38_cu100_pyt14.tar.bz2`, instead of simply `linux-64/pytorch3d-0.1.1-py38_cu100.tar.bz2`.
Also some tiny cleanup of circleci config.
Other alternatives: (1) forcing users to update pytorch and pytorch3d together, (2) trying to get away with one build for multiple pytorch versions.
Reviewed By: nikhilaravi
Differential Revision: D20599039
fbshipit-source-id: 20164eda4a5141afed47b3596e559950d796ffc9
Summary: Interface and working implementation of ragged KNN. Benchmarks (which aren't ragged) haven't slowed. New benchmark shows that ragged is faster than non-ragged of the same shape.
Reviewed By: jcjohnson
Differential Revision: D20696507
fbshipit-source-id: 21b80f71343a3475c8d3ee0ce2680f92f0fae4de
Summary: The conda build process generates some files of its own, which we don't want to catch in our test for copyright notices.
Reviewed By: nikhilaravi, patricklabatut
Differential Revision: D20868566
fbshipit-source-id: 76a786a3eb9a674d59e630cc06f346e8b82258a4
Summary:
Allows to initialize a Transform3D object with a batch of user-defined transformation matrices:
```
t = Transform3D(matrix=torch.randn(2, 4, 4))
```
Reviewed By: nikhilaravi
Differential Revision: D20693475
fbshipit-source-id: dccc49b2ca4c19a034844c63463953ba8f52c1bc
Summary:
1. Introduced weights to Umeyama implementation. This will be needed for weighted ePnP but is useful on its own.
2. Refactored to use the same code for the Pointclouds mask and passed weights.
3. Added test cases with random weights.
4. Fixed a bug in tests that calls the function with 0 points (fails randomly in Pytorch 1.3, will be fixed in the next release: https://github.com/pytorch/pytorch/issues/31421 ).
Reviewed By: gkioxari
Differential Revision: D20070293
fbshipit-source-id: e9f549507ef6dcaa0688a0f17342e6d7a9a4336c
Summary: Do not create output files with invalid inputs to `save_{obj,ply}()`.
Reviewed By: bottler
Differential Revision: D20720282
fbshipit-source-id: 3b611a10da6f6eecacab2a1900bf16f89e2000aa
Summary:
Similar to D20392526, PLY files without vertices or faces should be allowed:
- a PLY with only vertices can represent a point cloud
- a PLY without any vertex or face is just empty
- a PLY with faces referencing inexistent vertices has invalid data
Reviewed By: gkioxari
Differential Revision: D20400330
fbshipit-source-id: 35a5f072603fd221f382c7faad5f37c3e0b49bb1
Summary:
Enable `black` + `isort` (via `pyfmt`) i.e. `BLACK` fbsource linter.
NOTE: the `BLACK` fbsource linter (and `black` itself) is (by design) ***not*** configurable. This forces aligning the existing options used by the tools invoked in `dev/linter.sh` (for 3rd party developers) with `BLACK` fbsource linting. Without this reconciliation, the different linters (used internally or by 3rd party developers) would simply conflict with each other resulting in artificial back-and-forth changes (for instance line width which `BLACK` forces to 88 characters).
Reviewed By: nikhilaravi
Differential Revision: D20558374
fbshipit-source-id: 614fa00664f8eb9d2de7438c29b807dfbf36ad20
Summary: Run linter after recent changes. Fix long comment in knn.h which clang-format has reflowed badly. Add crude test that code doesn't call deprecated `.type()` or `.data()`.
Reviewed By: nikhilaravi
Differential Revision: D20692935
fbshipit-source-id: 28ce0308adae79a870cb41a810b7cf8744f41ab8
Summary:
OBJ files without vertices or faces should be allowed:
- an OBJ with only vertices can represent a point cloud
- an OBJ without any vertex or face is just empty
- an OBJ with faces referencing inexistent vertices has invalid data
Reviewed By: gkioxari
Differential Revision: D20392526
fbshipit-source-id: e72c846ff1e5787fb11d527af3fefa261f9eb0ee
Summary:
Implements K-Nearest Neighbors with C++ and CUDA versions.
KNN in CUDA is highly nontrivial. I've implemented a few different versions of the kernel, and we heuristically dispatch to different kernels based on the problem size. Some of the kernels rely on template specialization on either D or K, so we use template metaprogramming to compile specialized versions for ranges of D and K.
These kernels are up to 3x faster than our existing 1-nearest-neighbor kernels, so we should also consider swapping out `nn_points_idx` to use these kernels in the backend.
I've been working mostly on the CUDA kernels, and haven't converged on the correct Python API.
I still want to benchmark against FAISS to see how far away we are from their performance.
Reviewed By: bottler
Differential Revision: D19729286
fbshipit-source-id: 608ffbb7030c21fe4008f330522f4890f0c3c21a
Summary: CPU-only builds should be fixed by this change
Reviewed By: nikhilaravi
Differential Revision: D20598014
fbshipit-source-id: df098ec4c6c93d38515172805fe57cac7463c506
Summary: use assertClose in some tests, which enforces shape equality. Fixes some small problems, including graph_conv on an empty graph.
Reviewed By: nikhilaravi
Differential Revision: D20556912
fbshipit-source-id: 60a61eafe3c03ce0f6c9c1a842685708fb10ac5b
Summary: By accident I think, we have been nudging people to install python 3.6. Using 3.8 is fine.
Reviewed By: nikhilaravi
Differential Revision: D20597240
fbshipit-source-id: 7fb778f1b84746db28b6eef763564af5c5fffcd7
