GitLab

Previous weights are not compatible with current PyTorch

* Fix lint following #1695

* V2

* V3

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).

* scriptability checks

* tests upds

* linter upds

* linter upds

* upds

* tuple list changes

* linter updates

* Make R-CNN models less verbose in script mode

* Fix typo in warning message

* remove windows skips from video_utils tests, now that they pass

* replace lambda in videoclips in order to be pickled on windows and update tests

* tgz updates

* tgz updates

* tgz updates

* Update KeypointRCNN weights with correct file

* Fix model

* Fix

* add scriptable transform: center_crop

* add test: center_crop

* add scriptable transform: five_crop

* add scriptable transform: five_crop

* add scriptable transform: fix minor issues

The current link to the dataset is dead. The change links to the dataset's author's personal page, which describe the dataset and is also referenced at https://github.com/fyu/lsun.

* 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

* Add tests for results in script vs eager mode

This copies some logic from `test_jit.py` to check that a TorchScript'ed
model's outputs are the same as outputs from the model in eager mode.

To support differences in TorchScript / eager mode outputs, an
`unwrapper` function can be provided per-model.

* Fix inception, use PYTORCH_TEST_WITH_SLOW

* Update

* Remove assertNestedTensorObjectsEqual

* Add PYTORCH_TEST_WITH_SLOW to CircleCI config

* Add MaskRCNN unwrapper

* fix prec args

* Remove CI changes

* update

* Update

* remove expect changes

* Fix tolerance bug

* Fix breakages

* Fix quantized resnet

* Fix merge errors and simplify code

* DeepLabV3 has been fixed

* Temporarily disable jit compilation

* 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

* code changes to enable onnx export for keypoint rcnn

* add import

* fix copy paste error

* remove changes that induced BC

* Re-enable tests that have been disabled

* Remove outdated comment

* Remove outdated comment

Rename with_bias() to bias(), and output_channels() to out_channels() in C++ conv layer options usage (#1576)

* enable faster rcnn test

* flake8

* smaller image size

* set min/max

* Fix inconsistent NMS implementation

* Improve tests for NMS

* Remove unnecessary using statement

* Add Scriptable Transform: Grayscale

* add scriptable transforms: rgb_to_grayscale

* add scriptable transform: rgb_to_grayscale

* add scriptable transform: rgb_to_grayscale

* add scriptable transform: rgb_to_grayscale

* update code: rgb_to_grayscale

* add test: rgb_to_grayscale

* update parameters: rgb_to_grayscale

* add scriptable transform: rgb_to_grayscale

* update rgb_to_grayscale

* update rgb_to_grayscale

* quantizable googlenet

* Minor improvements

* Rename basic_conv2d with conv_block plus additional fixes

* More renamings and fixes

* Bugfix

* Fix missing import for mypy

* Add pretrained weights

* Unify video metadata in VideoClips

* Bugfix

* Make tests a bit more robust

* Improve readability of affine transformation code

* Make shear transformation area preserving

The previous shear implementation did not preserve area, and we
implement a version that does.

The formula used was verified with the following sympy code:

from sympy import Matrix, cos, sin, tan, simplify
from sympy.abc import x, y, phi

Xs = Matrix(
        [[1, -tan(x)],
         [0, 1]]
        )

Ys = Matrix(
        [[1, 0],
         [-tan(y), 1]]
        )

R = Matrix(
        [[cos(phi), -sin(phi)],
         [sin(phi), cos(phi)]]
        )

RSS = Matrix(
        [[cos(phi - y)/cos(y), -cos(phi - y)*tan(x)/cos(y) - sin(phi)],
         [sin(phi - y)/cos(y), -sin(phi - y)*tan(x)/cos(y) + cos(phi)]])

print(simplify(R * Ys * Xs - RSS))

One thing that is not clear (and could be tested) is whether avoiding
the explicit products and calculations in _get_inverse_affine_matrix
really gives performance benefits - compared to doing the explicit
calculation done in _test_transformation.

* Use np.matmul instead of @

The @ syntax is not supported in Python 2.