GitLab

* hello

* Try with DYLD_FALLBACK_LIBRARY_PATH
Co-authored-by: Edgar Andrés Margffoy Tuay <andfoy@gmail.com>

* WIP

* clang

* docs

* extracted out common utils

* Use better quantization function and pass tensors as parameters

* proper dequantization

* Some tests

* Dequantization optimization, seems to gain a few ms

* clang-format

* again

* more correct test. Had to remove optimization although it almost works

* Also test aligned=True

* remove useless part

* more docs and comments

* Put back optimization with more robust test

* Added check for index upper bound

* avoid possible overflow

* Move common function into common.h

* oops

* scale=1,zero_point=0 makes more sense

* Force batch size of 1 to prevent any indexingbug

* format

* format again

* updated docstring

* put back description comment for pre_calc_bilinear_interpolate

* revert most changes to docstring as it's taken care of in another PR

* Add quantized version of nms

* Added tests

* Compute areas only once

* remove calls to dequantize_val

* fix return type for empty tensor

* flake8

* remove use of scale as it gets cancelled out

* simpler int convertion in tests

* explicitly set ovr to double

* add tests for more values of scale and zero_point

* comment about underflow

* remove unnecessary accessor

* properly convert to float for division

* Add comments about underflow

* explicitely cast coordinates to float to allow vectorization

* clang

* clang  again

* hopefully OK now

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* new batched_nms implem

* flake8

* hopefully fix torchscipt tests

* Use where instead of nonzero

* Use same threshold (4k) for CPU and GPU

* Remove use of argsort

* use views again

* remove print

* trying stuff, I don't know what's going on

* previous passed onnx checks so the error isn't in _vanilla func. Trying to return vanilla now

* add tracing decorators

* cleanup

* wip

* ignore new path with ONNX

* use vanilla if tracing...????

* Remove script_if_tracing decorator as it was conflicting with _is_tracing

* flake8

* Improve coverage
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* Replace type T with accumulator.

* Upcast tensors of box ops to avoid overflow in multiplications.

Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* Speedup test_ucf101 (#2623 

* Speedup Cmake builds (#3186)

* Speedup test_autoaugment (#3190)

* Speedup DeformConvTester (#3191)

* Speedup InceptionV3 and GoogleNet on Windows (#3196)

* Replacing all torch.jit.annotations with typing

* Replacing remaining typing

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* Enable support for images without annotations

* Ensuring gradient propagates to RegressionHead.

* Rewriting losses to remove branching.

* Fix the seed on DeformConv autocast test.

* Add modulation input for DeformConv2D

* lint

* Patch for GPU CI

* Remove bad cache on CI

* Moving Autocast mtest in RoiOpTester.

* Adding autocast test on DeformConvTester.

* Adding autocast test on NMSTester.

* Fixing styles.

* Change default eps value of FrozenBN.

* Update the unit-tests.`

* Update the expected values.

* Revert the expected value and use original eps=0 value for flaky tests.

* Post init change of eps.

* Styles.

* feat: Updated FrozenBN eps to align with BatchNorm

* feat: Added eps to __repr__ of FrozenBN

* test: Updated unittest of __repr__ for FrozenBN

* test: Updated unittest for eps value in BN and FrozenBN

* fix: Revert FrozenBN eps value

* test: Revert test on eps alignment between FrozenBN and BN

* feat: Added __repr__ to MultiScaleRoIAlign

* test: Added unittest for __repr__ of MultiScaleRoIAlign

* feat: Added feature map names in __repr__

* test: Updated unittest
Co-authored-by: vfdev <vfdev.5@gmail.com>

* Adding checks on forward and backward passes.

* Adding unit-tests.

* fixies small bug in box_convert

* activates jit test

* Passes JIT test

* fixes typo

* adds error tests, removes assert

* implements to proposal2

* adds boxes conversion

* adds documentation

* adds xywh tests

* fixes small typo

* adds tests

* Remove sphinx theme

* corrects assertions

* cleans code as per suggestion
Signed-off-by: Aditya Oke <okeaditya315@gmail.com>

* reverts assertion

* fixes to assertEqual

* fixes inplace operations

* Adds docstrings

* added documentation

* changes tests

* moves code to box_convert

* adds more tests

* Apply suggestions from code review

Let's leave those changes to a separate PR

* fixes documentation
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* tries adding genaralized_iou

* fixes linting

* Adds docs for giou, iou and box area

* fixes lint

* removes docs to fixup in other PR

* linter fix

* Cleans comments

* Adds tests for box area, iou and giou

* typo fix for testCase

* fixes typo

* fixes box area test

* fixes implementation

* updates tests to tolerance

* [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

* 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>

* Add more tests to NMS

* Fix lint

* feat: Added eps argument to FrozenBatchNorm2d

* test: Added unittest for eps addition in FrozenBatchNorm2d

See #2169

* fix: Reverted forward changes for JIT fuser

* fix: Added back n argument for backward-compatibility

* fix: Fixed FrozenBatchNorm2d forward

Added back eps

* feat: Specified deprecation warnings in FrozenBatchNorm2d

* test: Added unittest for deprecation warninig in FrozenBatchNorm2d

* style: Fixed lint

* style: Fixed block comment lint

* feat: Restored support of tuple of Tensors for roi_align & roi_pool

* test: Added unittest for Tensor sequence support by region pooling

* test: Fixed typo in unittest

* test: Fixed data type

* test: Fixed roi pooling tensor unittest

* test: Fixed box format conversion

* feat: Added number of features in FrozenBatchNorm2d repr

While BatchNorm layers have extensive information in their repr, FrozenBatchNorm2d has one

* refactor: Refactored FrozenBatchNorm2d __repr__

* test: Added unittest for FrozenBatchNorm2d __repr__

* style: Removed blank lines in test_ops

* refactor: Avoids creating an extra attribute for __repr__

* style: Switched __repr__ to f-string

Since support of Python version ealier than 3.6 have been dropped, f-string can be used.

* fix: Fixed typo in __repr__

* style: Switched unittest .format to f-string

* improve stability of test_nms_cuda

This change addresses two issues:

_create_tensors_with_iou() creates test data for the NMS tests. It
takes care to ensure at least one pair of boxes (1st and last) have
IoU around the threshold for the test. However, the constructed
IoU for that pair is _so_ close to the threshold that rounding
differences (presumably) between CPU and CUDA implementations may
result in one suppressing a box in the pair and the other not.
Adjust the construction to ensure the IoU for the box pair is
near the threshold, but far-enough above that both implementations
should agree.

Where 2 boxes have nearly or exactly the same score, the CPU and
CUDA implementations may order them differently. Adjust
test_nms_cuda() to check only that the non-suppressed box lists
include the same members, without regard for ordering.

* adjust assertion in test_nms_cuda

The CPU and CUDA nms implementations each sort the box scores
as part of their work, but the sorts they use are not stable. So
boxes with the same score maybe be processed in opposite order
by the two implmentations.

Relax the assertion in test_nms_cuda (following the model in
pytorch's test_topk()) to allow the test to pass if the output
differences are caused by similarly-scored boxes.

* improve stability of test_nms_cuda

Adjust _create_tensors_with_iou() to ensure we create at least
one box just over threshold that should be suppressed.

* Add test for large batches in DeformConv2d

* Clean-up and (try) fix DeformConv2d

* Simplifications and bugfixes

* Try fix CUDA now

* remove sys.version_info == 2

* remove sys.version_info < 3

* remove from __future__ imports

* add checkout/assert in roi_pool

* add checkout/assert in roi_align

* move check_roi_boxes_shape func to ops/_utils.py

* add tests

* fix CI

* fix CI
Co-authored-by: Guanheng Zhang <zhangguanheng@devfair0197.h2.fair>

* 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

* Simlify and organize test_ops.

We perform the following:

- Simplify the functions slow_roi_pooling, slow_ps_roi_pooling, slow_ps_roi_align and bilinear_interpolate (including finding and removing a semi-bug in slow_ps_roi_pooling, which used bin_w instead of bin_h);
- Wrote a slow_roi_align function, that was missing;
- Create a base class testing all combinations of forward/backward, cpu/cuda, contiguous/non-contiguous;
- Organize all testing inside the base class with _test_forward and _test_backward (which can be easily overriden if a parciular op needs something different); an Op class then only needs to implement fn, get_script_fn, and expected_fn.

A few points:
- We are using the same inputs for all tests, and not trying all possible inputs in the domain of a given operation. One improvement would be to test more diverse inputs, and to personalize the inputs for some ops (e.g. different inputs for pooling ops and align ops).
- Running all tests is quite slow (~1 min only for CPU tests), so that can possibly be improved.

* Reduce input size used in gradcheck.

gradcheck can be quite costly, and it was causing OOM errors and making
the tests slow. By reducing the size of the input, the test speed is
down to 3 seconds for the CPU tests.

Other points:
- We remove an unused namedtuple;
- We inherit from object for better Python 2 compatibility;
- We remove a hardcoded pool_size from the TorchScript functions, and
add it as a parameter instead.

* Replace Tensor by torch.Tensor in type annotations.

This should fix lint errors.

* Fix inconsistent NMS implementation

* Improve tests for NMS

* Remove unnecessary using statement

Updated all raw asserts to corresponding unittest.TestCase.assert. See #1483