When I replace multiple maxtools with softtools, I find out Nan in tensor

在图像识别任务上，把最后分类池化层换为softpool时，识别精度由之前的90%降低到几乎为0，这有可能是什么原因造成的呢？

First of all, great work with SoftPool!

When I was adapting some code to test SoftPool for the first time, I left a stride=1 on the SoftPool2d constructor and noticed this inconsistent behavior on CPU (PyTorch implementation) vs GPU (CUDA implementation).

The shape of the output is different on the two implementations. With kernel 2 and stride 1 over a 3x3 input, I expected a 2x2 output, but the CUDA implementation gives 3x3. The example below illustrates well the difference.

In [1]: import torch

In [2]: from SoftPool import SoftPool2d

In [3]: x = torch.tensor([[[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]]])

In [4]: x
Out[4]: 
tensor([[[[1., 2., 3.],
          [4., 5., 6.],
          [7., 8., 9.]]]])

In [5]: SoftPool2d(kernel_size=2, stride=1)(x.cpu())
Out[5]: 
tensor([[[[4.5888, 5.5888],
          [7.5888, 8.5888]]]])

In [6]: SoftPool2d(kernel_size=2, stride=1)(x.cuda())
Out[6]: 
tensor([[[[1.0000, 1.7311, 2.7311],
          [3.8577, 4.5888, 5.5888],
          [6.8577, 7.5888, 8.5888]]]], device='cuda:0')

It looks to me that this is a bug on the way the CUDA path handles the output shape. On CUDA_SOFTPOOL2d, it does something like oH = H // stride[0] while it should do something like oH = (H - kernel_height) // stride[0] + 1. Of course some other parts of the code would need to be adapted as well.

Hello,

i was remain unsuccessful to install via following commonds:

$ git clone https://github.com/alexandrosstergiou/SoftPool.git
$ cd SoftPool-master/pytorch
$ make install

So, i have created and directly call SoftPool.py file. and it works fine. However, by commenting softpool_cuda.forward_2d(input.contiguous(), kernel, stride, output)

Because this library not properly installed, so i am unable to "import softpool_cuda", is there any trick to directly call these CUDA, and .cpp files in Python import? So, i can make softpool_cuda function work?

Waiting for reply...

The original pool operation can set the parameter Ceil_Mode =True, then how can the soft pool achieve the same effect, because there is a dimension mismatch problem in my replacement

Hi,I have trouble in "make install". The detail:FAILED: /pytorch/build/temp.linux-x86_64-3.6/CUDA/softpool_cuda_kernel.o
#pragma omp parallel for if ((end - begin) >= grain_size)
^
ninja: build stopped: subcommand failed.
python3.6/subprocess.py", line 438, in run
output=stdout, stderr=stderr)
subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1.
RuntimeError: Error compiling objects for extension
Makefile:2: recipe for target 'install' failed
make: *** [install] Error 1
Could you please help me?

Hi,thanks for your great idea!

I have a question in " softpool_cuda.forward_2d(input.contiguous(), kernel, stride, output) "，I just don't know what this line of code means ？What do you want to do with this line of code？

class CUDA_SOFTPOOL2d(Function):
    @staticmethod
    @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
    def forward(ctx, input, kernel=2, stride=None):
        # Create contiguous tensor (if tensor is not contiguous)
        no_batch = False
        if len(input.size()) == 3:
            no_batch = True
            input.unsqueeze_(0)
        B, C, H, W = input.size()
        kernel = _pair(kernel)
        if stride is None:
            stride = kernel
        else:
            stride = _pair(stride)
        oH = (H - kernel[0]) // stride[0] + 1
        oW = (W - kernel[1]) // stride[1] + 1
        output = input.new_zeros((B, C, oH, oW))
        softpool_cuda.forward_2d(input.contiguous(), kernel, stride, output)   #### what does this line mean ?####
        ctx.save_for_backward(input)
        ctx.kernel = kernel
        ctx.stride = stride
        if no_batch:
            return output.squeeze_(0)
        return output

Looking forward to your reply,thanks!

Thanks for your work!

There're some errors when installing.

`rm -rf *.egg-info
rm -rf build dist
python setup.py install
running install
running bdist_egg
running egg_info
creating SoftPool.egg-info
writing SoftPool.egg-info/PKG-INFO
writing dependency_links to SoftPool.egg-info/dependency_links.txt
writing top-level names to SoftPool.egg-info/top_level.txt
writing manifest file 'SoftPool.egg-info/SOURCES.txt'
writing manifest file 'SoftPool.egg-info/SOURCES.txt'
installing library code to build/bdist.linux-x86_64/egg
running install_lib
running build_py
creating build
creating build/lib.linux-x86_64-3.8
creating build/lib.linux-x86_64-3.8/SoftPool
copying SoftPool/idea.py -> build/lib.linux-x86_64-3.8/SoftPool
copying SoftPool/init.py -> build/lib.linux-x86_64-3.8/SoftPool
running build_ext
building 'softpool_cuda' extension
creating /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8
creating /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA
Emitting ninja build file /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/build.ninja...
Compiling objects...
Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
[1/2] :/usr/local/cuda-11.1:/usr/local/cuda-11.1/bin/nvcc --generate-dependencies-with-compile --dependency-output /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda_kernel.o.d -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/TH -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/THC -I:/usr/local/cuda-11.1:/usr/local/cuda-11.1/include -I/home/n504/anaconda3/envs/pytorch17/include/python3.8 -c -c /home/n504/Downloads/SoftPool/pytorch/CUDA/softpool_cuda_kernel.cu -o /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda_kernel.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options ''"'"'-fPIC'"'"'' -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="gcc"' '-DPYBIND11_STDLIB="libstdcpp"' '-DPYBIND11_BUILD_ABI="cxxabi1011"' -DTORCH_EXTENSION_NAME=softpool_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_86,code=sm_86 -std=c++14
FAILED: /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda_kernel.o
:/usr/local/cuda-11.1:/usr/local/cuda-11.1/bin/nvcc --generate-dependencies-with-compile --dependency-output /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda_kernel.o.d -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/TH -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/THC -I:/usr/local/cuda-11.1:/usr/local/cuda-11.1/include -I/home/n504/anaconda3/envs/pytorch17/include/python3.8 -c -c /home/n504/Downloads/SoftPool/pytorch/CUDA/softpool_cuda_kernel.cu -o /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda_kernel.o -D__CUDA_NO_HALF_OPERATORS -D__CUDA_NO_HALF_CONVERSIONS_ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options ''"'"'-fPIC'"'"'' -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1011"' -DTORCH_EXTENSION_NAME=softpool_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_86,code=sm_86 -std=c++14
/bin/sh: 1: :/usr/local/cuda-11.1:/usr/local/cuda-11.1/bin/nvcc: not found
[2/2] c++ -MMD -MF /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda.o.d -pthread -B /home/n504/anaconda3/envs/pytorch17/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/TH -I/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/THC -I:/usr/local/cuda-11.1:/usr/local/cuda-11.1/include -I/home/n504/anaconda3/envs/pytorch17/include/python3.8 -c -c /home/n504/Downloads/SoftPool/pytorch/CUDA/softpool_cuda.cpp -o /home/n504/Downloads/SoftPool/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda.o -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1011"' -DTORCH_EXTENSION_NAME=softpool_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++14
cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++
In file included from /home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/ATen/Parallel.h:140:0,
from /home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/csrc/api/include/torch/utils.h:3,
from /home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/csrc/api/include/torch/nn/cloneable.h:5,
from /home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/csrc/api/include/torch/nn.h:3,
from /home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/csrc/api/include/torch/all.h:13,
from /home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/torch/extension.h:4,
from /home/n504/Downloads/SoftPool/pytorch/CUDA/softpool_cuda.cpp:1:
/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/include/ATen/ParallelOpenMP.h:83:0: warning: ignoring #pragma omp parallel [-Wunknown-pragmas]
#pragma omp parallel for if ((end - begin) >= grain_size)

ninja: build stopped: subcommand failed.
Traceback (most recent call last):
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/utils/cpp_extension.py", line 1667, in _run_ninja_build
subprocess.run(
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/subprocess.py", line 512, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1.

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
File "setup.py", line 4, in
setup(
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/init.py", line 153, in setup
return distutils.core.setup(**attrs)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/core.py", line 148, in setup
dist.run_commands()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/dist.py", line 966, in run_commands
self.run_command(cmd)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/dist.py", line 985, in run_command
cmd_obj.run()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/command/install.py", line 67, in run
self.do_egg_install()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/command/install.py", line 109, in do_egg_install
self.run_command('bdist_egg')
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/cmd.py", line 313, in run_command
self.distribution.run_command(command)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/dist.py", line 985, in run_command
cmd_obj.run()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/command/bdist_egg.py", line 167, in run
cmd = self.call_command('install_lib', warn_dir=0)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/command/bdist_egg.py", line 153, in call_command
self.run_command(cmdname)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/cmd.py", line 313, in run_command
self.distribution.run_command(command)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/dist.py", line 985, in run_command
cmd_obj.run()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/command/install_lib.py", line 11, in run
self.build()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/command/install_lib.py", line 107, in build
self.run_command('build_ext')
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/cmd.py", line 313, in run_command
self.distribution.run_command(command)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/dist.py", line 985, in run_command
cmd_obj.run()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/command/build_ext.py", line 79, in run
_build_ext.run(self)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/Cython/Distutils/old_build_ext.py", line 186, in run
_build_ext.build_ext.run(self)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/command/build_ext.py", line 340, in run
self.build_extensions()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/utils/cpp_extension.py", line 708, in build_extensions
build_ext.build_extensions(self)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/Cython/Distutils/old_build_ext.py", line 195, in build_extensions
_build_ext.build_ext.build_extensions(self)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/command/build_ext.py", line 449, in build_extensions
self._build_extensions_serial()
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/command/build_ext.py", line 474, in _build_extensions_serial
self.build_extension(ext)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/setuptools/command/build_ext.py", line 196, in build_extension
_build_ext.build_extension(self, ext)
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/distutils/command/build_ext.py", line 528, in build_extension
objects = self.compiler.compile(sources,
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/utils/cpp_extension.py", line 529, in unix_wrap_ninja_compile
_write_ninja_file_and_compile_objects(
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/utils/cpp_extension.py", line 1354, in _write_ninja_file_and_compile_objects
_run_ninja_build(
File "/home/n504/anaconda3/envs/pytorch17/lib/python3.8/site-packages/torch/utils/cpp_extension.py", line 1683, in _run_ninja_build
raise RuntimeError(message) from e
RuntimeError: Error compiling objects for extension
Makefile:2: recipe for target 'install' failed
make: *** [install] Error 1`

And I'm sure that I've cuda installed and added it into path.
output of command $nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2020 NVIDIA Corporation Built on Mon_Oct_12_20:09:46_PDT_2020 Cuda compilation tools, release 11.1, V11.1.105 Build cuda_11.1.TC455_06.29190527_0
output of command $echo $PATH:
/home/n504/anaconda3/envs/pytorch17/bin:/home/n504/anaconda3/condabin:/usr/local/cuda-11.1/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin:/usr/local/cuda-11.1/bin

hi， ‘make install’ is succeed，but running ‘make test’ occurs some errors.
`cd test-files && python test.py
--- Initial checks for forward ---

Checking 1D CPU ...
PASSED
Checking 1D GPU ...
cudaCheckError() failed : invalid device function
Segmentation fault (core dumped)
Makefile:9: recipe for target 'test' failed
make: *** [test] Error 139
`

hello
This is my mistake!
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [1, 256, 40, 40]], which is output 0 of ReluBackward0, is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

The error ends here！
grad_input = torch.zeros_like(ctx.saved_tensors[0])
May I ask what caused the problem
thank you！

Import softtool_ cuda error，his package is softpool in CUDA folder_ cuda.cpp？

Hello, when I run make install, I get the following error:


thank you

Hi,
I installed the package on Colab with the commands you provide. The installation process seemed successful, but when I tried to import soft_pool2d, I got the following error message:

ModuleNotFoundError                       Traceback (most recent call last)
<ipython-input-2-89f2a448160d> in <module>()
----> 1 from SoftPool import soft_pool2d, SoftPool2d

1 frames
/content/SoftPool/pytorch/SoftPool/idea.py in <module>()
      5 from torch.nn.modules.utils import _triple, _pair, _single
      6 
----> 7 import softpool_cuda
      8 
      9 

ModuleNotFoundError: No module named 'softpool_cuda'

Is there any way to use SoftPool on Colab? Thanks in advance.

when I input "softpool" in my own CNN, this bug happens

RuntimeError: output_grad.is_contiguous() INTERNAL ASSERT FAILED at "E:\SoftPool-master\pytorch\CUDA\softpool_cuda.cpp":119, please report a bug to PyTorch. output_grad must be a contiguous tensor

Hey. I am getting a few errors during the installation process.
Into the SoftPool/pytorch folder, I run make install as explained in the README file, and during the process, three errors show up in the console repeatedly:

CUDA/softpool_cuda_kernel.cu(252): error: type name is not allowed
CUDA/softpool_cuda_kernel.cu(252): error: expected an expression
CUDA/softpool_cuda_kernel.cu(252): error: a pointer to a bound function may only be used to call the function

And then:

Error limit reached.
100 errors detected in the compilation of "CUDA/softpool_cuda_kernel.cu".
Compilation terminated.
error: command '/usr/local/cuda-11.3/bin/nvcc' failed with exit status 1
Makefile:2: recipe for target 'install' failed
make: *** [install] Error 1

For quick inspection, here are the lines that are causing the error (at least the first 100 of them). It looks like the issue is with AT_DISPATCH_FLOATING_TYPES_AND_HALF:

I have tried different CUDA versions (9.0, 10.2, 11.3) and different GCC/G++ versions (5, 8 and 9), but no success. I am using Ubuntu 16.04 with an RTX 2080 Ti.

$ gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/lib/gcc/x86_64-linux-gnu/9/lto-wrapper
OFFLOAD_TARGET_NAMES=nvptx-none:hsa
OFFLOAD_TARGET_DEFAULT=1
Target: x86_64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Ubuntu 9.3.0-23ubuntu1~16.04' --with-bugurl=file:///usr/share/doc/gcc-9/README.Bugs --enable-languages=c,ada,c++,go,brig,d,fortran,objc,obj-c++,gm2 --prefix=/usr --with-gcc-major-version-only --program-suffix=-9 --program-prefix=x86_64-linux-gnu- --enable-shared --enable-linker-build-id --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --libdir=/usr/lib --enable-nls --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --with-default-libstdcxx-abi=new --enable-gnu-unique-object --disable-vtable-verify --enable-plugin --with-system-zlib --with-target-system-zlib=auto --enable-objc-gc=auto --enable-multiarch --disable-werror --with-arch-32=i686 --with-abi=m64 --with-multilib-list=m32,m64,mx32 --enable-multilib --with-tune=generic --enable-offload-targets=nvptx-none=/build/gcc-9-oatUX7/gcc-9-9.3.0/debian/tmp-nvptx/usr,hsa --without-cuda-driver --enable-checking=release --build=x86_64-linux-gnu --host=x86_64-linux-gnu --target=x86_64-linux-gnu
Thread model: posix
gcc version 9.3.0 (Ubuntu 9.3.0-23ubuntu1~16.04)

Thanks!

I am a rookie. I have the following installation problems when using Windows system

I want to use sfotpool instead of convolution in YOLO to complete image downsampling, but I don’t know whether to use 1D, 2D or 3D. Can you give me some suggestions?

RuntimeError: input.is_contiguous() INTERNAL ASSERT FAILED at "/export/w/SoftPool/pytorch/CUDA/softpool_cuda.cpp":120, please report a bug to PyTorch. input must be a contiguous tensor

Can soft pool be used as global soft pooling which can replace global average pooling?

Thank you for your sharing~
However I have some doubts,

1. this code represents the weight value is calculation by using channel dimension on the same spatial position, rather than applying spatial region?
   

   SoftPool/pytorch/SoftPool/idea.py

   Line 181 in 7cd12c7

   e_x = torch.sum(torch.exp(x),dim=1,keepdim=True)

2. for the last pooling layer, global avg pooling is better than softpool?

Looking forward to your reply

Any comparison with LIP on several benchmarks?

when I load your pre-trained model, it seems to have some error.

"state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)"
"load_state_dict_from_url" does not define.

I download the model resnet-50_best.pth, and load it but error occurs.
Error(s) in loading state_dict for ResNet:
Missing key(s) in state_dict: "conv1.weight", "bn1.weight", "bn1.bias", "bn1.running_mean", "bn1.running_var", "layer1.0.conv1.weight", "layer1.0.bn1.weight", "layer1.0.bn1.bias", "layer1.0.bn1.running_mean", "layer1.0.bn1.running_var", ..........

Can give me some advice to load your pre-trained model. It would be nice if you could provide detailed code. Thanks.

provide ImageNet weight only change maxpool to softpool?

have any differnet from torchvision.models?

I wonder why there is no parameter 'padding' in softpool function. When the kernel size can not be divided by input size, what will you do? And what should I modify so that i can use padding.

Hi, If I set force_inplace to False, CUDA will report an error:
RuntimeError: CUDA error: an illegal memory access was encountered.
cudaCheckError() failed : an illegal memory access was encountered.
Could you tell me why please?

Hi, I find that you didn't release pre-trained inception model. Do you have any plans to release the model?

I have successfully installed the SoftPool in my computer, I have tested your train.py code on Resnet50 and it is normal without errors.
But when I use the softpool2d in my own code , I want to use the softpool on a bs×C×125×160 tensor to get a bs×C×1×20 tensor, so I set the kernel=(125,8), stride=(125,8)(i.e.,self.pool=SoftPool2d(kernel_size=(125,8),stride=(125,8))). In backward step, I meet the error:
'SoftPool-master/pytorch/CUDA/softpool_cuda.cpp":119, please report a bug to PyTorch. output_grad must be a contiguous tensor'.
Could you help to solve my problem? My pytorch version is 1.6.

Why are there obvious afterimages in the video processing results?

It's OK to change “nn.MaxPool2d(2)“ to “out = soft_pool2d(out, 2)”, but changing ”nn.MaxPool2d((2, 1))” to ”out = soft_pool2d(out, (2,1)) “ or "out = soft_pool2d(out, 2,1)"will result in wrong dimension. What's the reason?

i want to add a parameter about "padding",i wonder if it is oK to do that ?

Excuse me, I am a frashman,
I find that "softpool" no padding, and when i do softpool, it is wrong, can you explain it?
Sorry about my ENG! Hope you can understand my means ^-^ Thank you!

hi, thanks for your great work.
Could you provide model weight of action recognition that trained on Kinectics 700 dataset?

have no padding?

Traceback (most recent call last):
File "/media/mm/62f65e0e-b396-44ac-b1aa-b2fb260e70d1/mm/experiment/DE-Resnet18/main.py", line 251, in
main(args, logger)
File "/media/mm/62f65e0e-b396-44ac-b1aa-b2fb260e70d1/mm/experiment/DE-Resnet18/main.py", line 171, in main
train(model, train_loader, optimizer, criterion, epoch, args, logger)
File "/media/mm/62f65e0e-b396-44ac-b1aa-b2fb260e70d1/mm/experiment/DE-Resnet18/main.py", line 61, in train
loss.backward()
File "/home/mm/anaconda3/lib/python3.7/site-packages/torch/tensor.py", line 221, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "/home/mm/anaconda3/lib/python3.7/site-packages/torch/autograd/init.py", line 132, in backward
allow_unreachable=True) # allow_unreachable flag
File "/home/mm/anaconda3/lib/python3.7/site-packages/torch/autograd/function.py", line 89, in apply
return self._forward_cls.backward(self, *args) # type: ignore
File "/media/mm/62f65e0e-b396-44ac-b1aa-b2fb260e70d1/mm/experiment/DE-Resnet18/softpool.py", line 45, in backward
saved[-1][torch.isnan(saved[-1])] = 0

My environment info is:
GPU: Titan
CUDA: 10.1.243
cuDNN: 7.6.5
PyTorch: 1.7.1

Hi Alex,

I just got a problem for installation. The error is: 'RuntimeError: Error compiling objects for extension'. I am using pytorch 1,7 and cuda 10.1. I do not know if it is related to my Pytorch or CUDA version. Do you have any advice?

Best wishes,
Qing

File "/home/csu/anaconda3/envs/wsy/lib/python3.6/site-packages/SoftPool-1.1-py3.6-linux-x86_64.egg/SoftPool/idea.py", line 192, in soft_pool2d
if torch.isnan(x).any():
RuntimeError: CUDA error: an illegal memory access was encountered
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.

The changes in my model are as follows：
def init(self, num_classes: int = 1000) -> None:
super(AlexNet, self).init()
self.conv1 = nn.Sequential(
......
)

self.conv2 = nn.Sequential(
    nn.Conv2d(64, 192, kernel_size=5, padding=2),
    nn.ReLU(inplace=True),
    SoftPool2d(kernel_size=3, stride=2), # This is where I changed
)

When I replace the softpool of the self.conv1 structure, no error occurs at runtime. However, an error occurred while replacing the self.conv2 structure.
The GPU I use is RTX 3090. The version of pytorch I use is 1.9.0.

Hi， thank you for your great work.
Could you please provide the keras implement？

hi, I have some trouble when i install softpool. The detail is that im follow "Installation" and use "make install",then im encountered the mistake:

(mos3d) root@LUCAS-DEV-01a694:~/MotionSeg3D/SoftPool/pytorch# make install
rm -rf *.egg-info
rm -rf build dist
python setup.py install
running install
/usr/local/envs/mos3d/lib/python3.7/site-packages/setuptools/command/install.py:37: SetuptoolsDeprecationWarning: setup.py install is deprecated. Use build and pip and other standards-based tools.
setuptools.SetuptoolsDeprecationWarning,
/usr/local/envs/mos3d/lib/python3.7/site-packages/setuptools/command/easy_install.py:147: EasyInstallDeprecationWarning: easy_install command is deprecated. Use build and pip and other standards-based tools.
EasyInstallDeprecationWarning,
running bdist_egg
running egg_info
creating SoftPool.egg-info
writing SoftPool.egg-info/PKG-INFO
writing dependency_links to SoftPool.egg-info/dependency_links.txt
writing top-level names to SoftPool.egg-info/top_level.txt
writing manifest file 'SoftPool.egg-info/SOURCES.txt'
reading manifest file 'SoftPool.egg-info/SOURCES.txt'
writing manifest file 'SoftPool.egg-info/SOURCES.txt'
installing library code to build/bdist.linux-x86_64/egg
running install_lib
running build_py
creating build
creating build/lib.linux-x86_64-3.7
creating build/lib.linux-x86_64-3.7/SoftPool
copying SoftPool/init.py -> build/lib.linux-x86_64-3.7/SoftPool
copying SoftPool/idea.py -> build/lib.linux-x86_64-3.7/SoftPool
running build_ext
building 'softpool_cuda' extension
creating build/temp.linux-x86_64-3.7
creating build/temp.linux-x86_64-3.7/CUDA
gcc -pthread -B /usr/local/envs/mos3d/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include/torch/csrc/api/include -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include/TH -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include/THC -I/usr/local/cuda-10.2/include -I/usr/local/envs/mos3d/include/python3.7m -c CUDA/softpool_cuda.cpp -o build/temp.linux-x86_64-3.7/CUDA/softpool_cuda.o -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE="gcc" -DPYBIND11_STDLIB="libstdcpp" -DPYBIND11_BUILD_ABI="cxxabi1011" -DTORCH_EXTENSION_NAME=softpool_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++14
cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++
/usr/local/cuda-10.2/bin/nvcc -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include/torch/csrc/api/include -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include/TH -I/usr/local/envs/mos3d/lib/python3.7/site-packages/torch/include/THC -I/usr/local/cuda-10.2/include -I/usr/local/envs/mos3d/include/python3.7m -c CUDA/softpool_cuda_kernel.cu -o build/temp.linux-x86_64-3.7/CUDA/softpool_cuda_kernel.o -D__CUDA_NO_HALF_OPERATORS -D__CUDA_NO_HALF_CONVERSIONS_ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options '-fPIC' -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE="_gcc" -DPYBIND11_STDLIB="_libstdcpp" -DPYBIND11_BUILD_ABI="_cxxabi1011" -DTORCH_EXTENSION_NAME=softpool_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_70,code=compute_70 -gencode=arch=compute_70,code=sm_70 -std=c++14
CUDA/limits.cuh(119): error: identifier "inf" is undefined

CUDA/limits.cuh(120): error: identifier "inf" is undefined

CUDA/limits.cuh(128): error: identifier "inf" is undefined

CUDA/limits.cuh(129): error: identifier "inf" is undefined

4 errors detected in the compilation of "/tmp/tmpxft_00280a89_00000000-6_softpool_cuda_kernel.cpp1.ii".
error: command '/usr/local/cuda-10.2/bin/nvcc' failed with exit code 1
Makefile:2: recipe for target 'install' failed
make: *** [install] Error 1

my environment ：
pytorch：1.11.0
cuda:10.2
gcc:7.5

The project could not compile successfully with cuda-9.0, what should I do to change the code cuda/softpool_cuda.h?

#define CHECK_CUDA(x) AT_ASSERT(x.is_cuda(), #x " must be a CUDA tensor") #define CHECK_CONTIGUOUS(x) AT_ASSERT(x.is_contiguous(), #x " must be a contiguous tensor");

I already tried TORCH_INTERNAL_ASSERT, it did not work. Thanks alot.

The log is as follows:

cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++ CUDA/softpool_cuda.cpp:53:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp:53:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:54:23: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(output); ^ CUDA/softpool_cuda.cpp:54:23: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:73:28: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(output_grad); ^ CUDA/softpool_cuda.cpp:73:28: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:74:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp:74:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:75:27: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input_grad); ^ CUDA/softpool_cuda.cpp:75:27: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:95:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp:95:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:96:23: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(output); ^ CUDA/softpool_cuda.cpp:96:23: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:119:28: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(output_grad); ^ CUDA/softpool_cuda.cpp:119:28: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:120:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp:120:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:121:27: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input_grad); ^ CUDA/softpool_cuda.cpp:121:27: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:146:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp:146:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:147:23: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(output); ^ CUDA/softpool_cuda.cpp:147:23: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:175:28: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(output_grad); ^ CUDA/softpool_cuda.cpp:175:28: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:176:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp:176:22: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp:177:27: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CHECK_INPUT(input_grad); ^ CUDA/softpool_cuda.cpp:177:27: error: macro "AT_ASSERT" passed 2 arguments, but takes just 1 CUDA/softpool_cuda.cpp: In function ‘int softpool1d_forward_cuda(at::Tensor, std::tuple<int>, std::tuple<int>, at::Tensor)’: CUDA/softpool_cuda.cpp:47:23: error: ‘AT_ASSERT’ was not declared in this scope #define CHECK_CUDA(x) AT_ASSERT(x.is_cuda(), #x " must be a CUDA tensor") ^ CUDA/softpool_cuda.cpp:49:24: note: in expansion of macro ‘CHECK_CUDA’ #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x); ^ CUDA/softpool_cuda.cpp:53:5: note: in expansion of macro ‘CHECK_INPUT’ CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp: In function ‘int softpool1d_backward_cuda(at::Tensor, at::Tensor, std::tuple<int>, std::tuple<int>, at::Tensor)’: CUDA/softpool_cuda.cpp:47:23: error: ‘AT_ASSERT’ was not declared in this scope #define CHECK_CUDA(x) AT_ASSERT(x.is_cuda(), #x " must be a CUDA tensor") ^ CUDA/softpool_cuda.cpp:49:24: note: in expansion of macro ‘CHECK_CUDA’ #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x); ^ CUDA/softpool_cuda.cpp:73:5: note: in expansion of macro ‘CHECK_INPUT’ CHECK_INPUT(output_grad); ^ CUDA/softpool_cuda.cpp: In function ‘int softpool2d_forward_cuda(at::Tensor, std::tuple<int, int>, std::tuple<int, int>, at::Tensor)’: CUDA/softpool_cuda.cpp:47:23: error: ‘AT_ASSERT’ was not declared in this scope #define CHECK_CUDA(x) AT_ASSERT(x.is_cuda(), #x " must be a CUDA tensor") ^ CUDA/softpool_cuda.cpp:49:24: note: in expansion of macro ‘CHECK_CUDA’ #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x); ^ CUDA/softpool_cuda.cpp:95:5: note: in expansion of macro ‘CHECK_INPUT’ CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp: In function ‘int softpool2d_backward_cuda(at::Tensor, at::Tensor, std::tuple<int, int>, std::tuple<int, int>, at::Tensor)’: CUDA/softpool_cuda.cpp:47:23: error: ‘AT_ASSERT’ was not declared in this scope #define CHECK_CUDA(x) AT_ASSERT(x.is_cuda(), #x " must be a CUDA tensor") ^ CUDA/softpool_cuda.cpp:49:24: note: in expansion of macro ‘CHECK_CUDA’ #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x); ^ CUDA/softpool_cuda.cpp:119:5: note: in expansion of macro ‘CHECK_INPUT’ CHECK_INPUT(output_grad); ^ CUDA/softpool_cuda.cpp: In function ‘int softpool3d_forward_cuda(at::Tensor, std::tuple<int, int, int>, std::tuple<int, int, int>, at::Tensor)’: CUDA/softpool_cuda.cpp:47:23: error: ‘AT_ASSERT’ was not declared in this scope #define CHECK_CUDA(x) AT_ASSERT(x.is_cuda(), #x " must be a CUDA tensor") ^ CUDA/softpool_cuda.cpp:49:24: note: in expansion of macro ‘CHECK_CUDA’ #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x); ^ CUDA/softpool_cuda.cpp:146:5: note: in expansion of macro ‘CHECK_INPUT’ CHECK_INPUT(input); ^ CUDA/softpool_cuda.cpp: In function ‘int softpool3d_backward_cuda(at::Tensor, at::Tensor, std::tuple<int, int, int>, std::tuple<int, int, int>, at::Tensor)’: CUDA/softpool_cuda.cpp:47:23: error: ‘AT_ASSERT’ was not declared in this scope #define CHECK_CUDA(x) AT_ASSERT(x.is_cuda(), #x " must be a CUDA tensor") ^ CUDA/softpool_cuda.cpp:49:24: note: in expansion of macro ‘CHECK_CUDA’ #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x); ^ CUDA/softpool_cuda.cpp:175:5: note: in expansion of macro ‘CHECK_INPUT’ CHECK_INPUT(output_grad); ^ error: command 'gcc' failed with exit status 1 Makefile:2: recipe for target 'install' failed make: *** [install] Error 1

I found an interesting result in your paper. The speed of sum operation is slower compared with the average operation by a huge margin. Could you please explain the reason for this phenomenon? Thank you!

Thank you for your excellent work. But I have a question.
I trained the model with GPU, and the SoftPool2d parameter force_inplace was set to the default value False.
But the results I measured using GPU and CPU in the test are different, and the mAP of CPU is much lower than that of GPU. And if I change the force_inplace=True in the test, the mAP of the GPU is equal to that of the CPU, which is much lower than the previous GPU（force_inplace=False）

(base) cx@amax:~/CL/HashNet-master/SoftPool-master/pytorch$ make install
rm -rf *.egg-info
rm -rf build dist
python setup.py install
Traceback (most recent call last):
File "setup.py", line 2, in
from torch.utils.cpp_extension import BuildExtension, CUDAExtension
ModuleNotFoundError: No module named 'torch'
Makefile:2: recipe for target 'install' failed
make: *** [install] Error 1

what's matter with that ?

Hello, I reported the following error when installing the 3090 graphics card
FAILED: /data/home/lwl/GaitSet_experiment/SoftPool-master/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda_kernel.o /usr/local/cuda/bin/nvcc -I/data/home/lwl/anaconda3/envs/exp38/lib/python3.8/site-packages/torch/include -I/data/home/lwl/anaconda3/envs/exp38/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -I/data/home/lwl/anaconda3/envs/exp38/lib/python3.8/site-packages/torch/include/TH -I/data/home/lwl/anaconda3/envs/exp38/lib/python3.8/site-packages/torch/include/THC -I/usr/local/cuda/include -I/data/home/lwl/anaconda3/envs/exp38/include/python3.8 -c -c /data/home/lwl/GaitSet_experiment/SoftPool-master/pytorch/CUDA/softpool_cuda_kernel.cu -o /data/home/lwl/GaitSet_experiment/SoftPool-master/pytorch/build/temp.linux-x86_64-3.8/CUDA/softpool_cuda_kernel.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options ''"'"'-fPIC'"'"'' -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1011"' -DTORCH_EXTENSION_NAME=softpool_cuda -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_86,code=sm_86 -std=c++14 nvcc fatal : Unsupported gpu architecture 'compute_86'

hello，I am a novice，how to install windows system without make

Hi, thank you for your great idea. It's very interesting.

Because efficientnet get a lot of great results in Imagenet.
I just want to know have you ever implemented your idea in efficientnet before?
(Because efficientnet get a lot of great results in Imagenet.)

If you had done it before, What's the result?

Thank you again for everything you've done.

RuntimeError: input.is_contiguous() INTERNAL ASSERT FAILED at "CUDA/softpool_cuda.cpp":95, please report a bug to PyTorch. input must be a contiguous tensor.
I met the above error, how to solve it? Thanks.

Hi, how to solve the image padding operation?