Added grid_sample backward batch rule #284
Conversation
Description: - Added grid_sample backward batch rule: CPU and CUDA - Updated tests Notes: I had to expand on dim 0 in most of the cases and could not use tricks like in forward pass when batch dim is merged either with channel or H_out due to wrong grid grads in these cases
vfdev-5
force-pushed
the
add-grid_sample-bw
branch
from
d500cb9
to
42ebfac
Compare
| VMAP_SUPPORT("grid_sampler_3d_backward", GRID_SAMPLE_BW_BATCH_RULE(grid_sampler_3d_backward)); | ||
| VMAP_SUPPORT("cudnn_grid_sampler_backward", CUDNN_GRID_SAMPLE_BW_BATCH_RULE(cudnn_grid_sampler_backward)); |
There was a problem hiding this comment.
Do both these get exercised in the tests?
There was a problem hiding this comment.
Yes, cudnn_grid_sampler_backward is tested by test/test_ops.py::TestOperatorsCUDA::test_vmapvjp_has_batch_rule_nn_functional_grid_sample_cuda_float32
and grid_sampler_3d_backward is tested by one of test/test_ops.py::TestOperatorsCPU::test_vmapvjp_has_batch_rule_nn_functional_grid_sample_cpu_float32
|
Hmm, tests are failing with build issue but I'm not sure that they're related |
|
Test is failing because we're getting an unexpected success from I'm also not sure if this is related but just to flag if it is |
|
There is a cuda test failing at build due to pytorch 1.7.1 installation by conda (somehow), https://app.circleci.com/pipelines/github/pytorch/functorch/798/workflows/e842f064-bc17-4925-b330-88f0aa0c94a1/jobs/4776?invite=true#step-103-333 |
Just reran the tests (sorry that it reran all of them...I mean to only do that one). It looks like that just hit a weird race condition or something funky happened with the checkout since the root errors were a bunch of "file not found"s. I think the CUDA test failures should be similar across the 3 versions after this |
|
I synced the PR to the |
) * Added grid_sample backward batch rule Description: - Added grid_sample backward batch rule: CPU and CUDA - Updated tests Notes: I had to expand on dim 0 in most of the cases and could not use tricks like in forward pass when batch dim is merged either with channel or H_out due to wrong grid grads in these cases * Code updates according to the review * Updated OutOfPlacePlumbing.cpp to the latest pytorch
) * Added grid_sample backward batch rule Description: - Added grid_sample backward batch rule: CPU and CUDA - Updated tests Notes: I had to expand on dim 0 in most of the cases and could not use tricks like in forward pass when batch dim is merged either with channel or H_out due to wrong grid grads in these cases * Code updates according to the review * Updated OutOfPlacePlumbing.cpp to the latest pytorch
Description:
std::arrayin signatureNotes:
I had to expand on dim 0 in most of the cases and could not use
tricks like in forward pass when batch dim is merged either with channel or H_out
due to wrong grid grads in these cases
Related to #240