acecoooool / roialign-roipool-pytorch

C++ extension implementation of RoIAlign & RolPool (both GPU and CPU) for PyTorch

RoI-op-pytorch

C++ extension of RoIPool & RoIAlign (both CPU and GPU) in PyTorch，this code is converted from caffe2 operation. （need pytorch 0.4.0）

**Warning：**You may change AT_CHECK to AT_ASSERT（0.4 version using AT_ASSERT， and latest version using AT_CHECK）

**Note: **

1. roi_xxx_cpu.cpp&roi_xxx_binding.cpp：contains the cpu version of forward and backward operation.(Note: roi_xxx_binding.cpp is for pybind, you can put this code into roi_xxx_cpu.cpp as well)
2. roi_xxx_kernel.cu&roi_xxx_cuda.cpp：contains the cuda version of forward and backward operation.
3. main.py&temp.h&CMakeLists.txt：help you to debug in c++ code, rather than to run python setup.py install to debug. (Note: only support cpu version ~ I don't know how to debug .cu code 😣)
4. setup.py：you can run python setup.py install to install this operation as a package (You can find this package in you python site-package)
5. roi_xxx.py：wrap .cpp code to pytorch's Function & Module ，there is also a small demo testing.

Install

cd roixxx   # roipool or roialign
python setup.py install

RoI Pooling

The "strategy" of roi-pooling in this implementaion likes the follow picture:（:joy: so bad picture）

Note: (please stand on point view rather than block view)

1. scale=0.5
2. dotted line is the range of "seleted area" (int form in [left, right) and [top, bottom))

RoI Align

Note: left sample=1, right sample=2

There are several good resource to explain these two operations:

• Region of interest pooling explained
• ROI Align --- chinese
• ROI Align --- youtube

Reference

1. caffe2 operator：most of the code is coming from here.
2. extension-cpp: tutorial
3. detectorch

roialign-roipool-pytorch's Issues

Question about batch data.

Hello,
First of all, really thank you for your hard work and sharing.

The reason why I write the issue is that I don't know exactly a meaning of "batch index" (the first element of "rois")

I hope to use your roi_align (or roi_pooling) to batch tensor. In that case, how to use your function?

Thank you

What are the requirements of this project?

my pytorch version is 1.8.1 and I meet this problem when running python setup.py install

In file included from roi_align_binding.cpp:2:0:
roi_align_cpu.cpp:241:57: 错误：‘rois_col’在此作用域中尚未声明
                        {channel, h, w}, pool_h, pool_w, rois_col, sample, grad_in.data<float>());
                                                         ^~~~~~~~
roi_align_cpu.cpp:241:57: 附注：suggested alternative: ‘rois’
                        {channel, h, w}, pool_h, pool_w, rois_col, sample, grad_in.data<float>());
                                                         ^~~~~~~~
                                                         rois
roi_align_cpu.cpp:241:88: 错误：expected primary-expression before ‘float’
                        {channel, h, w}, pool_h, pool_w, rois_col, sample, grad_in.data<float>());
                                                                                        ^~~~~

CUDA Eerror:an illegal memory access was encountered

class ROIPoolFunction(Function): @staticmethod def forward(ctx, feat, rois, pool_h, pool_w, scale, train): ctx.rois = rois ctx.feat_size = feat.size() ctx.pool_h = pool_h ctx.pool_w = pool_w if train: ctx.memory = torch.zeros((rois.size(0), feat.size(1), pool_h, pool_w), dtype=torch.int) else: ctx.memory = torch.zeros(0) if feat.is_cuda: ctx.memory = ctx.memory.cuda() output = roi_pool_cuda.forward_cuda(feat, rois, pool_h, pool_w, scale, ctx.memory) else: output = roi_pool_cpu.forward_cpu(feat, rois, pool_h, pool_w, scale, ctx.memory) return output

ctx.memory = torch.zeros(0)
i found this line will lead to the runtime error :CUDA Eerror:an illegal memory access was encountered

Awesome work, but there exists some bugs!

Thanks for your awesome work! I have learned a lot and fixed the bugs.

Origin boxes

Why's origin boxes = (1.6 1.6 9.2 11.0)? Also, how to calculate bin_w, bin_h and scale_boxes?

Question about how to debug in pytorch0.4.1

In pytorch0.4.1, there's no libATen.so
I'm new to write CMakeList
So, do you have any suggestion about how can I write CMakeList to debug in pytorch0.4.1?

there seems to be a mistake?

Thank you for the code.But there seems to be a mistake?
in RoIAlign-RoIPool-pytorch/roialign/roi_align_kernel.cu , Line 85：

it should be :
const T *offset_input = input + (roi_batch_idx * channel+ c) * h * w;

Question about grad_out of roi_pooling

Thanks for your code. I have a question, why you add
grad_out = grad_out.contiguous() if not grad_out.is_contiguous() else grad_out in the roi_pool.py L32? Thanks.