GitLab

* refactoring disco backend code

* removed get_psi as member function and instead put it in _disco_convolution

* setting seeds in tests more consistently

* parametrized test classes to ensure that tests are always run on both CPU and GPU (if available)

* cleaning up

https://github.com/NVIDIA/torch-harmonics/compare/main...azrael417:torch-harmonics:tkurth/mauro-rebase

Also match the gradient output to the input, in terms of memory layout

Detect memory layout (B,C,H,W) (stride for C should be 1, if not, fix it)

This ensures that the backwards kernel is fast

putting qy in shared is a little faster

Changing internal memory layout means we can leave code in standard shape and
only change layout external to kernel

fixing bug in quadrature weights for full attention. Adding better unit tests for attention. Cleanup in the cuda code.

* small hotfix for lobatto grid precomputation routine

* adding lobatto grid to tests

* adding caching

* replacing many numpy calls with torch calls

* bumping up version number to 0.7.6

switched psi tensor computation to double precision and implemented a fudge factor for theta_cutoff to avoid aliasing issues with the grid width

* added analytic gradients to the gradient_analysis notebook

* fixing sht unittest to not check the roundtrip gradient but sht and isht individually

* Updated changelog

* changing distributeds SHT to use dim=-3 as the channel dimension for distributed transpose

* Fixed formatting in tests

* Adding bash script for running test suite

* Replaced asserts regarding number of dimensions in tensor with checks

* missed one instance of python3

* fused multiplication of quadrature into multiplication of the Psi tensor

* Added quadrature change to changelog

* removing persistent quad weights tensor as it isn't needed anymore

* added pretty-prenting for convolution modules

* adjusting default value in convolution test

* adding cuda kernels for disco conv

* making psi_idx an attribute

* adding license headers

* adding author files

* reorganizing files

* draft implementation

* added conditional installation to setup.py

* formatting changes

* removing triton kernel in DISCO convolution

* updated github actions

* updated Readme and changelog

* adding another guard for the cuda installation

* renaming the  cuda extension

* simplifying setup.py

* minor bugfix

* Bbonev/cuda disco cleanup (#32)

* cleanup of disco convolutions based on CUDA extension

* fixing unittest

* changing version to experimental 0.7.0a

* initial rewrite of the distributed convolution with CUDA

* fixing streams

* need to fix install options

* fixing streams

* undid setup.py changes

* reset setup.py

* including CUDAStream

* adjusted the precomputation of theta_cutoff. If you rely on this, your models will not be backwards-compatible.

* adjusting theta_cutoff in the unittest

* adding newly refactored kernels for faster compile

* Tkurth/cuda disco distributed fix (#34)

* attempt to make disco distributed

* working distributed convolutions

* fixing distributed conv

* working distributed disco

* removing irrelevant extra argument

* using stream functions from at instead of c10

* using stream functions from at instead of c10, small fix

* Bbonev/disc even filters (#35)

* initial working commit with new convention of counting collocation points across the diameter instead of across the radius

* fixed a bug in the computation of the even kernels

* changing heuristic for computing theta_cutoff

* Fixing unittest

* Readability improvements

* reworked normalization of filter basis functions

* implemented discrete normalization of disco filters

* relaxing tolerances in convolution unit test

* bugfix to correctly support unequal scale factors in latitudes and longitudes

* hotfix to a bug in the imports

* Bbonev/distributed disco refactor (#37)

* cleaned up normalization code in convolution

* formatting changes in distributed convolution

* Fixing default theta_cutoff to be the same in distributed and local case

* fixed distributed convolution to support the same normalization as non-distributed one

* readability improvements

* fixed initial scale of convolution parameter weights and fixed naming of the normalization routine

* Updated Readme.md

* added comment in Dockerfile regarding older architectures

---------
Co-authored-by: Thorsten Kurth <tkurth@nvidia.com>
Co-authored-by: Boris Bonev <bbonev@nvidia.com>

* initial implementation of distributed DISCO layer

* working distributed convolution

* working refactored serial conv transpose with torch kernel

* working distributed conv and transposed conv when using the python kernel

* working distributed convolution with torch kernel

* fixed triton kernel tests

* adding print statement to debug CI

* adjusting tolerances in local convolution unittest

---------
Co-authored-by: Boris Bonev <bbonev@nvidia.com>

Changed the code to only implicitly use sparse tensors in the modules, in order to enable compatibility with DDP

* Moved convolutions and exposed them directly

* Added transposition to the unit test

* Minor bugfix in CPU version of DISCO transpose code

* Adding convolution tests to CI

* Added gradient check

* Checking the weight grad as well

* Added test for anisotropic kernels

* Fixing the precomputation of the Psi matrix

* moving tests to a parent folder

* removing tqdm

* deleting deprecated distributed tests

* Moving distributed test

* adapting workfolow

* Added some more comments to make the code more understandable

* more detailed explanation for the derivation of the rotation angles

* added another comment

---------
Co-authored-by: Thorsten Kurth <tkurth@nvidia.com>

* working distributed fwd SHT with flexible sharding and without padding

* fixing distributed sht with new logic

* fixed distributed SHT and ISHT with flexible padding

* working unittest for distributed fwd SHT

* distributed tests converted to unit tests

* bumping version number up to 0.6.4

* fixing small splitting bug in th distributed

* updated changeloc, removed tests folder