GitLab

Operator registration of functions with noexcept doesn't work on some compilers, see https://github.com/pytorch/pytorch/issues/48667

* Removing python headers and code for Python2 compatibility.

* Restoring python3 inits.

* Add modulation input for DeformConv2D

* lint

* Patch for GPU CI

* Remove bad cache on CI

* Fixing types.

* Dispatcher + Autocast.

* + Autograd.

* Formating.

* Clean up and refactor PSROIPool implementation:
- Remove primitive const declaration from method names.
- Using references when possible.
- Fix variable naming.

* Restore include headers.

* New line at end of file.

* Resolving conflict, final cleanup, ordering method consistently across files.

* Fixing types.

* Dispatcher + Autocast.

* + Autograd.

* Formating.

* Fixing return casting with autocast.

* Clean up and refactor ROIPool implementation:
- Remove primitive const declaration from method names.
- Using references when possible.

* Restore include headers.

* New line at end of file.

* Fixing types.

* Dispatcher + Autocast.

* + Autograd.

* Clean up and refactor PSROIAlign implementation:
- Remove primitive const declaration from method names.
- Using references when possible.
- Sync naming of internal methods with other ops.

* Restoring names of internal methods to avoid conflicts.

* Restore include headers.

* Splitting tuples of stride, padding and dilation of DeformConv.

* Fixing types.

* Dispatcher + Autocast.

* + Autograd.

* Moving contiguous() convertions away dispatcher and into the implementations.

* Removing rvalue references.

* Ensure torchvision operators are registered in C++ via weak symbols

* Add note to README on how to ensure that torchvision operators are available in C++

* Fix dllimport/dllexport on windows, format files

* Factor out common macros in single file

* Expose cuda_version in the API, use it to avoid pruning of ops initializer

* add autocasting on ROCm

* enable ROIAlign autocasting on ROCm

* enable NMS autocasting on ROCm

* fix to use correct torch CUDA APIs

* Fixes Xiao's repro

* Ports nms to use full dispatcher

* Move HIPGuard to nms_cuda

* clang-format

* run models in test_models.py on GPU if available

* Francisco's comment, also disable cuda model tests to see if CPU alone still passes

* cuda tests now pass locally, although still not comparing to saved numerics

* add note for thing to ask francisco

* Allow cuda and cpu tests to share a data file

* ignore suffix if unneeded

* Skip autocast numerics checks for a few models

* Add roi_align test
Co-authored-by: Michael Carilli <mcarilli@nvidia.com>

* Switch torchvision registrations to new operator registration API.

This is still registering everything as catchalls, so we're really just
moving deck chairs around, but payoff is coming soon.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Port roi_align to actually use dispatcher
Signed-off-by: Edward Z. Yang <ezyang@fb.com>

* Added code to support creating extension on ROCm

* max -> fmaxf conversion for hipification

* added WITH_HIP flag for hipExtension

* added appropriate headers for HIP build

* use USE_ROCM in condition to build

* change fmaxf and fminf calls

* fminf -> min

* fix the check for ROCM_HOME

* more robust checking for rocm pytorch

* add check for pytorch version before using HIP extensions

* conditional reading of ROCM_HOME

* Aligned flag in the interfaces

* Aligned flag in the impl, and remove unused comments

* Handling empty bin in forward

* Remove raise error in roi_width

* Aligned flag in the Testcodes

* Add Deformable Convolution operation.

This adds the deformable convolution operation, as described in Deformable Convolutional Networks (https://arxiv.org/abs/1703.06211).

- The code is based on https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp ; the whole code was modified and refactored to remove redundancies and increase clarity, and to adapt it to torchvision.

- The CPU part is a direct copy of the CUDA code; it might make sense to do follow-up adjustments in the CPU code to simplify it / optimize it, or to reuse functionality between CPU and CUDA..

- We also add tests (with a non-trivial set of parameters); they can be made more robust by randomizing the parameters and executing multiple times.

* Update DeformConv to be more consistent w/ Conv2d

* rename some variables and arguments to match Conv2d;
* add optional bias;
* add weight, offset and bias as module parameters;
* remove the n_parallel_imgs parameter;
* Fix __repr__;
* etc..

Initialization of weight and bias is the same as in Conv2d, and
initialization of offsets to zero is the same as in the paper.

This also includes some other small unrelated fixes/improvements.

* Apply clang-format in DeformConv files.

* Import Optional type annotation

* Remove offset param from DeformConv2d module

- We pass the offset in the forward of DeformConv2d, instead of having
an internal parameter. This adds some complexity to creating the module
(e.g. now you have to worry about the output size, to create the
offset), but it gives more flexibility.
- We also use make_tuple for tuple creation, in an attempt to fix error
w/ older compilers.

* Replace abs by std::abs

Old gcc versions were giving wrong results here, because they would
resolve abs as int -> int, thus causing undesired truncation. Replacing
abs by std::abs should allow for correct overloading of abs as float -> float.

* Reorder declarations for clarity

* Reorder weight and offset args in deform_conv2d

We place offset arg before the weight arg, to be more
consistent with DeformConv2d.forward(input, offset)

* Replace abs by std::abs in DeformConv_cuda

* almost working...

* respond to comments

* add empty tensor op, handle different output types in generalized rcnn

* clean ups

* address comments

* more changes

* it's working!

* torchscript bugs

* add script/ eager test

* eval script model

* fix flake

* division import

* py2 compat

* update test, fix arange bug

* import division statement

* fix linter

* fixes

* changes needed for JIT master

* cleanups

* remove imagelist_to

* requested changes

* Make FPN backwards-compatible and torchscript compatible

We remove support for feature channels=0, but support for it was already a bit limited

* Fix ONNX regression

* added PSRoiAlign and PSRoiPool with C++ autograd and torch ops

* fixed linter errors

* fixed linter errors 2

* fixed linter errors 3

* Remove C++ extensions in favor of custom ops

* Remove unused custom_ops.cpp file

* Rename _custom_ops.py

* Reorganize functions

* Minor improvements and fixes

* Fix lint

* Fully scriptable ops

* Import types used by annotations

* Revert "Revert "Register Torchvision Ops as Cutom Ops (#1267)" (#1316)"

This reverts commit fe234fc8.

* Make import of C++ extensions lazy

* define python initialization functions for extension

* Fix lint

This reverts commit 78f169b5.

* Register torchvision ops

* install ORT only with python 3

* remane lib + address other comments

* fix lint

* fix lib copy

* find file with pattern instead of suffix

* use relative path

* revert rename and use imp to find lib

* fix typo

* Initial layout for layers with cpp extensions

* Move files around

* Fix import after move

* Add support for multiple types to ROIAlign

* Different organization

CUDA extensions work now

* Cleanups

* Reduce memory requirements for backwards

* Replace runtime_error by AT_ERROR

* Add nms test

* Add support for compilation using CPP extensions

* Change folder structure

* Add ROIPool cuda

* Cleanups

* Add roi_pool.py

* Fix lint

* Add initial structures folder for bounding boxes

* Assertion macros compatible with pytorch master (#540)

* Support for ROI Pooling (#592)

* ROI Pooling with tests. Fix for cuda context in ROI Align.

* renamed bottom and top to follow torch conventions

* remove .type().tensor() calls in favor of the new approach to tensor initialization (#626)

* Consistent naming for rois variable (#627)

* remove .type().tensor() calls in favor of the new approach to tensor initialization

* Consistent naming for rois variable in ROIPool

* ROIPool: Support for all datatypes (#632)

* Use of torch7 naming scheme for ROIAlign forward and backward

* use common cuda helpers in ROIAlign

* use .options() in favor of .type() where applicable

* Added tests for forward pass of ROIAlign, as well as more consistent naming scheme for CPU vs CUDA

* working ROIAlign cuda backwards pass

* working ROIAlign backwards pass for CPU

* added relevant headers for ROIAlign backwards

* tests for ROIAlign layer

* replace .type() with .options() for tensor initialization in ROIAlign layers

* support for Half types in ROIAlign

* gradcheck tests for ROIAlign

* updated ROIPool on CPU to work with all datatypes

* updated and cleaned tests for ROI Pooling

* Fix rebase problem

* Remove structures folder

* Improve cleanup and bugfix in test_layers

* Update C++ headers

* Add CUDAGuard to cu files

* Add more checks to layers

* Add CUDA NMS and tests

* Add multi-type support for NMS CUDA

* Avoid using THCudaMalloc

* Add clang-format and reformat c++ code

* Remove THC includes

* Rename layers to ops

* Add documentation and rename functions

* Improve the documentation a bit

* Fix some lint errors

* Fix remaining lint inssues

* Area computation doesn't add +1 in NMS

* Update CI to use PyTorch nightly

* Make NMS return indices sorted according to the score

* Address reviewer comments

* Lint fixes

* Improve doc for roi_align and roi_pool

* move to xenial

* Fix bug pointed by @lopuhin

* Fix RoIPool reference implementation in Python 2

Also fixes a bug in the clip_boxes_to_image -- this function needs a test!

* Remove change in .travis