GitLab

* Encapsulate and standardize deform_conv2d (#3074)

* Rename files.

* Standardizing method names.

* Adding anonymous namespaces.

* Applying C++ naming rules and alinging variable names across headers and cpp files.

* Syncing names across implementations.

* Rename deform_conv2d.h to deform_conv2d.cpp

* Use header files:
- Create header files for kernel implementation and remove definitions from vision_*.h files.
- Eliminate unnecessary headers and ensure all cpp include their headers.

* Change the naming convention for kernel implementations.

* Remove the _param postfix from the variables and standardizing names.

* Exposing public forward/backward methods to the C++ API and moving methods around to minimize git blame changes.

* Encapsulate and standardize nms (#3081)

* Syncing, where possible, the names of functions across devices.

* Adding all internal functions in anonymous namespaces.

* Renaming C++/CUDA kernel files and moving operator code from header to cpp file.

* Create foreach cpp file a separate header file with "public" functions.

* Removing unnecessary repeated includes.

* Update CMakeLists.txt to include all headers.

* Encapsulate and standardize ps_roi_align (#3082)

* Renaming C++ files & methods according to recommended naming conventions and aligning them with Python's API.
Syncing, where possible, the names of functions across devices.

* Adding all internal functions in anonymous namespaces.

* Renaming C++/CUDA kernel files and moving operator code from header to cpp file.

* Create foreach cpp file a separate header file with "public" functions.

* Removing unnecessary repeated includes.

* Encapsulate and standardize ps_roi_pool (#3084)

* Renaming C++ files & methods according to recommended naming conventions and aligning them with Python's API.

* Adding all internal functions in anonymous namespaces.

* Renaming C++/CUDA kernel files and moving operator code from header to cpp file.

* Create foreach cpp file a separate header file with "public" functions.

* Removing unnecessary repeated includes.

* Encapsulate and standardize roi_align (#3085)

* Renaming C++ files & methods according to recommended naming conventions and aligning them with Python's API.

* Adding all internal functions in anonymous namespaces.

* Renaming C++/CUDA kernel files and moving operator code from header to cpp file.

* Create foreach cpp file a separate header file with "public" functions.

* Removing unnecessary repeated includes.

* Encapsulate and standardize roi_pool  (#3088)

* Renaming C++ files & methods according to recommended naming conventions and aligning them with Python's API.

* Adding all internal functions in anonymous namespaces.

* Syncing variable names between the cpp files and their header files.

* Renaming C++/CUDA kernel files and moving operator code from header to cpp file.

* Create foreach cpp file a separate header file with "public" functions.

* Removing unnecessary repeated includes.

* Encapsulate and standardize new_empty_tensor_op (#3089)

* Renaming C++ files & methods according to recommended naming conventions and aligning them with Python's API.

* Create foreach cpp file a separate header file with "public" functions.

* Adding all internal functions in anonymous namespaces.

* Convert to const ref all possible parameters.

* Removing unnecessary repeated includes.

* Encapsulate and standardize C++ Ops - Clean up (#3094)

* Removing unnecessary repeated includes.

* Remove unnecessary vision_cpu.h, vision_cuda.h, autocast.h.

* Fixing naming convention and correcting method names on macros.

* Turn on clang formatter for cu files and fixing broken styles.

* Replace "#ifndef ... #define ... #endif" with "#pragma once" on header files.

* Adding operator methods in vision::ops namespace. (#3096)

* Adding operator methods in vision::ops namespace.

* Replace general.h with macros.h

* Adding vision.h to the necessary cpp files.

* 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.

* Clean up and refactor ROIAlign implementation:
- Remove primitive const declaration from method names.
- Passing as const ref instead of value where possible.
- Remove unnecessary headers.

* Adding back include for cpu.

* Restore include headers.

* 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.

* Clean up and refactor DeformConv implementation:
- Remove primitive const declaration from method names.
- Passing as const ref instead of value where possible.
- Aligning method names between cpu and cuda.

* Adding newline.

* Adding back include for cpu.

* Restoring method names of private methods to avoid conflicts.

* Restore include headers.

* Clean up and refactor ROIAlign implementation:
- Remove primitive const declaration from method names.
- Remove unnecessary headers.
- Aligning method names between cpu and cuda.

* Adding back include for cpu.

* Restoring method names of private 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.

- Remove unnecessary return.

* Convert to const reference and eliminate unnecessary bool casting.

* Removing unnecessary namespace use.

* 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

* Added rois shape check in C++

* Fixes code formatting

* Remove accidental include

* - Updated code according to the review
- Replaced old AT_ASSERT/ERROR by new TORCH_CHECK

* Adding checks on forward and backward passes.

* Adding unit-tests.

* Add CMake build pipeline

* Add CMake build workflow

* Add executable permissions to script

* Install cmake on Windows/MacOS

* Install conda-build before setting up MSVC

* Install PyTorch from nightly

* Do not use conda-build variables

* Add path to CMake

* Install libpng and libjpeg

* Perform make

* Call msbuild on Windows

* Add missing yq

* Use vc_env_helper

* Use string instruction

* Escape configuration option

* Remove configuration flag

* Try to pass -p

* Use caret to escape equal sign

* Escape string option in Windows

* Try to call other bat

* Remove Windows/GPU CMake

* Add tracing cpp test

* Script model instead of tracing it

* Try to register operators manually

* Use manylinux-cuda102

* Activate conda env on Linux

* Build and run sample tracing test

* Add empty echo

* Remove unnecessary register

* Copy headers on Mac

* Revert to 2xlarge

* Include /usr/local/include on Mac

* Install pillow on Windows

* Install future

* Install torchvision on Windows

* Set include flag

* Add torchlib to PATH

* Normalize path via cygpath

* Register ops on Windows

* Minor error correction

* Register CPU/GPU ops on DLL library and register ops via reference

* Install dataclasses

* Install dataclasses using pip

* Address clang formatting issue

* Try to use an actual GPU instance on Linux

* Remove extra environment section

* Declare environment explicitly

* Regenerate

* Pass env variables to Dokcer

* Regenerate circleci

* Test tracing on GPU

* Use GPU medium

* Regenerate

* Use cuda101

* Regenerate

* Do not use pre-trained weights

Avoids having to download pretrained files, which could cause flaky tests
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* add autocasting on ROCm

* enable ROIAlign autocasting on ROCm

* enable NMS autocasting on ROCm

* fix to use correct torch CUDA APIs

* [WIP] Fixes #2598
- Adjusted num_kernels and batch_size according to kMaxGridNum * CUDA_NUM_THREADS
- Tests to add

* - Redefined kMaxGridNum as max grid according to current CUDA device
- Added test to check the code from issue and compared grads CPU/CUDA

* Fixed static kMaxGridNum evaluation to dynamic

* Fix wrong clamping in RoIAlign with aligned=True

* Fix silly mistake

* Bugfix pointed out during code-review

* 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>

Fixes https://github.com/pytorch/vision/issues/2214#issuecomment-628636663

I don't know why the building is not working with `--expt-relaxed-constexpr` flag set, but it is generally a good idea to declare this as `__host__ __device__`

* Don't include CUDAApplyUtils.cuh

* fix format

* fix atomic

* fix the use of contiguous() in kernels

* clang-format

* add a contiguous in nms
Co-authored-by: Yuxin Wu <ppwwyyxx@users.noreply.github.com>

Torchvision includes at least 3 bits of code that calculate
box Intersection over Union values (and usually compare to
a threshold):

- box_iou in torchvision/ops/boxes.py
- devIoU in torchvision/csrc/cuda/nms_cuda.cu
- nms_cpu_kernel in torchvision/csrc/cpu/nms_cpu.cpp

The calculations were performed slightly differently between
those, leading to occasional differences in results.

Update devIoU to use the same method as the others for better
consistency.

This change improves agreement between the CPU and CUDA
calculations but the results can still differ slightly.
Setting NVCC_FLAGS to include "--fmad=true" would provide
still better agreement, but with likely cost to performance.

* Replace **.is_cuda() to just is_cuda()

* Replace type to scalar_type

* Fix lint, clang-format

* Fix lint, clang-format

* Add test for large batches in DeformConv2d

* Clean-up and (try) fix DeformConv2d

* Simplifications and bugfixes

* Try fix CUDA now

* fix shape error for deform conv gpu op

recover shape of columns for next iteration in for loops, previous version will cause error when batch_sz / n_parallel_imgs > 1

* fix shape error for deform conv cpu op

recover shape of columns for next iteration in for loops, previous version will cause error when batch_sz / n_parallel_imgs > 1

* 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

1. Let the IOU function compare with threshold. This avoid a division. Similar strategy is also used in https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/non_max_suppression_op.cu.cc
2. Only compute the upper triangle of the mask.

This speeds up the kernel about 20% (tested on GTX 1080Ti, with 20 input cases dumped from a Mask R-CNN inference job).

* 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

* Fix inconsistent NMS implementation

* Improve tests for NMS

* Remove unnecessary using statement

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

* fixed linter errors

* fixed linter errors 2

* fixed linter errors 3

* Use Tensor.data_ptr instead of .data

* use pytorch-nightly in CI

* #944
  MSBuild Compile time casting Error

* #944
 MSBuild Error  static_cast<Long> to static_cast<int64_t>

* Add eval.py
 Not Work find_contours

* Remove unnecessary file

* Lint