matteo-ronchetti / torch-radon

Computational Tomography in PyTorch

TorchRadon: Fast Differentiable Routines for Computed Tomography

TorchRadon is a PyTorch extension written in CUDA that implements differentiable routines for solving computed tomography (CT) reconstruction problems.

The library is designed to help researchers working on CT problems to combine deep learning and model-based approaches.

Main features:

• Forward projections, back projections and shearlet transforms are differentiable and integrated with PyTorch .backward().
• Up to 125x faster than Astra Toolbox.
• Batch operations: fully exploit the power of modern GPUs by processing multiple images in parallel.
• Transparent API: all operations are seamlessly integrated with PyTorch, gradients can be computed using .backward(), half precision can be used with Nvidia AMP.
• Half precision: storing data in half precision allows to get sensible speedups when doing Radon forward and backward projections with a very small accuracy loss.

Implemented operations:

• Parallel Beam projections
• Fan Beam projections
• Shearlet transform

Installation

Currently only Linux is supported, if you are running a different OS please use Google Colab or the Docker image.

Precompiled packages

If you are running Linux you can install Torch Radon by running:

wget -qO- https://raw.githubusercontent.com/matteo-ronchetti/torch-radon/master/auto_install.py  | python -

Google Colab

You can try the library from your browser using Google Colab, you can find an example notebook here.

Docker Image

Docker images with PyTorch CUDA and Torch Radon are available here.

docker pull matteoronchetti/torch-radon

To use the GPU in docker you need to use nvidia-docker

Build from source

You need to have CUDA and PyTorch installed, then run:

git clone https://github.com/matteo-ronchetti/torch-radon.git
cd torch-radon
python setup.py install

If you encounter any problem please contact the author or open an issue.

Benchmarks

The library is noticeably faster than the Astra Toolbox, especially when data is already on the GPU. Main disadvantage of Astra is that it only takes inputs which are on the CPU, this makes training end-to-end neural networks very inefficient. The following benchmark compares the speed of Astra Toolbox and Torch Radon:

If we set clip_to_circle=True (consider only the part of the image that is inside the circle) the speed difference is even larger:

These results hold also on a cheap laptop GPU:

Cite

If you are using TorchRadon in your research, please cite the following paper:

@article{torch_radon,
Author = {Matteo Ronchetti},
Title = {TorchRadon: Fast Differentiable Routines for Computed Tomography},
Year = {2020},
Eprint = {arXiv:2009.14788},
journal={arXiv preprint arXiv:2009.14788},
}

Testing

Install testing dependencies with pip install -r test_requirements.txt then test with:

nosetests tests/

torch-radon's Issues

how to import it when downloaded in anaconda

I have downloaded it in my anaconda environment, and I can see "torch-radon 1.0.0 pypi_0 pypi" if I do conda list.but there is wrong when i code"import torch-radon".How can i import it?

detector size

Hi! First of all thanks for the this great contribution.
I am using your torch_radon on a data which size are 972x972. I do the forward and I expect the respective sinogram size to be [angles, 972*sqrt(2)]. Why running your function the second dimension is 972 instead of sqrt(2)*972?
Is it possible to fix this?

benchmarks pyro-nn

are there any plans to benchmark it against pyro-nn

Small type typo in parameter_classes.h (v2.0.0)

Hi!

First of all, I'd like to appreciate the neat work @matteo-ronchetti has been doing on Torch-Radon!

I wanted to report a small typo that arises a compilation error when building Torch-Radon v2.0.0. In parameter_classes.h, both get_block_dim and get_grid_size headers return a wrong type inside ExecCfg class. I believe they should return vec3 instead of dim3.

Regards.

Backend FFT functions are not differentiable

The functions cuda_backend.rfft and cuda_backend.irfft don't implement a backward pass, so their gradient can't be computed. This makes BaseRadon.filter_sinogram unusable for network trainings.

Is there a reason why you don't use torch.fft.rfft and torch.fft.irfft?

pytorch 1.8.1 has a problm: module ‘torch‘ has no attribute "rfft"

pytorch 1.8.1 has a problm: module ‘torch‘ has no attribute "rfft"

error in pip install setup.py

Using CUDA_HOME=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.3
Compiling src\forward.template
C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.3/bin/nvcc -std=c++11 -ccbin=g++ -Xcompiler -fPIC -Xcompiler -static -Xcompiler -static-libgcc -Xcompiler -static-libstdc++ -I./include -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -DNDEBUG -O3 --generate-line-info --compiler-options -Wall -c src\forward.cu -o objs/cuda/forward.o
'C:/Program' 不是内部或外部命令，也不是可运行的程序
或批处理文件。
ERROR IN COMPILATION

Handle non-squared images

I'd like to have the possibility to compute the Radon Transform of non-squared images.
This is extremely useful in a number of imaging applications.

Publication

Hello,
do you have a publication for this package that we could cite?
Best,
Florian

torch.rfft not available anymore

Hi,
with pytorch >= 1.8 there is no function torch.rfft anymore, can you integrate an alternative?
Thanks!

Status of version 2

Hi,
I was wondering about version 2 of this library. Do you have any plans about when you are going to release it?
Currently, I am having troubles with the torch.fft modules of version 1, but the v2 branch does not build for me.
Any hints would be appreciated. Thanks for your help!

How to install the 3D projection in the v2 version, does it need to be installed in the same environment as the 2D?

Windows not supported?

The setup.py script leads me to believe the installation cannot be done on Windows, is that true?

Thanks

Colab updated to Python 3.10

As mentioned in the title, Colab updated to Python 3.10 and the Notebook you linked to does not work anymore. In particular, the autocomplier fails since only Python versions 3.6 3.7 and 3.8 are supported.

CUDA error: no kernel image is available for excecution on the device

I am receiving the following error when running the simple fbp example when caclulating the sinogram:

RuntimeError: CUDA error: no kernel image is available for execution on the device

My setup:
Nvidia GPU A6000
Pytorch 1.11.0 py3.9 cuda11.3 cudnn8.2.0.0
Cudatoolkit 11.3.1
Cudatoolkit-dev 11.4.1 - needed it to build TorchRadon from source

Torchradon was build without issues. Simple PyTorch calculations do not produce this error.

make: *** No rule to make target `install'. Stop.

Hi, I encounter a problem of "make: *** No rule to make target `install'. Stop." when I used "make install" .

Could you please help me to solve it. Thanks a lot. Hope Receive you reply soon.

AttributeError: 'RadonFanbeam' object has no attribute 'noise_generator'

Got error:
AttributeError: 'RadonFanbeam' object has no attribute 'noise_generator'
though I don't call noise_generator

Parallel Radon slightly off for simple example

Good work on this interesting package. I built 306ae46 and ran the following simple test where I expect that the Radon operator should be perfect at angles 0 and PI/2.

It seems that possibly the angle specification is wrong or there is some other bias in the geometry for the rays. Was wondering if there is an explanation for this and if it could be fixed.

import torch
from torch_radon import Radon

def test_simple_integrals(image_size=16):
    """Check that the forward radon operator works correctly at 0 and PI/2.

    When we project at angles 0 and PI/2, the foward operator should be the
    same as taking the sum over the object array along each axis.
    """
    angles = torch.tensor(
        [0.0, -np.pi / 2, np.pi, np.pi / 2],
        dtype=torch.float32,
        device='cuda',
    )
    radon = Radon(
        resolution=image_size,
        angles=angles,
        det_spacing=1.0,
        det_count=image_size,
        clip_to_circle=False,
    )

    original = torch.zeros(
        image_size,
        image_size,
        dtype=torch.float32,
        device='cuda',
    )
    original[image_size // 4, :] += 1

    data = radon.forward(original)
    data0 = torch.sum(original, axis=0)
    data1 = torch.sum(original, axis=1)

    print('\n', data[0].cpu().numpy())
    print(data0.cpu().numpy())
    print('\n', data[1].cpu().numpy())
    print(data1.cpu().numpy())
    print('\n', data[2].cpu().numpy())
    print(data0.cpu().numpy()[::-1])
    print('\n', data[3].cpu().numpy())
    print(data1.cpu().numpy()[::-1])

tests/test_parallel_beam.py::test_simple_integrals 
 [0.99632365 0.99632365 0.99632365 0.99632365 0.99632365 0.99632365
 0.99632365 0.99632365 0.99632365 0.99632365 0.99632365 0.99632365
 0.99632365 0.99632365 0.99632365 0.99632365]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]

 [ 0.        0.        0.        0.       16.000002  0.        0.
  0.        0.        0.        0.        0.        0.        0.
  0.        0.      ]
[ 0.  0.  0.  0. 16.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.]

 [0.99632365 0.99632365 0.99632365 0.99632365 0.99632365 0.99632365
 0.99632365 0.99632365 0.99632365 0.99632365 0.99632365 0.99632365
 0.99632365 0.99632365 0.99632365 0.99632365]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]

 [ 0.        0.        0.        0.        0.        0.        0.
  0.        0.        0.        0.       16.000002  0.        0.
  0.        0.      ]
[ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0. 16.  0.  0.  0.  0.]

When can torch-radon support RTX 4090?

I have a question, when can torch-radon support RTX 4090? RTX 4090 seems to be compute_capabilities=8.9.
thanks a lot!

Error of python setup.py install

Thank you for your fascinating codes.
I wonder, there may be a mistake in init.py (in torch-radon/build_tools file direcotory). When I run python setup.py install in linux, bash raises a mistake "UnboundLocalError: local variable 'compute_capabilities' referenced before assignment"

I found "compute_capabilities" is wrong in the format of "compute_capabilites" in two different lines n init.py (in torch-radon/build_tools file direcotory). After I fix it, the code runs smoothly.
Looking forward to checking in your free time!

windows error in 'python setup.py install'

Thank you for your code. My computer environment is Python 3.7.5, Cuda 10.1, and PyTorch 1.6. During installation, I encountered an issue where I couldn't find the specified file, and the error code is shown in the image. May I ask how to solve this problem?

Will accept Pull Requests on my fork / recent builds available on conda-forge

If you're reading this, it's because you're wondering if this package is still active. No. The last updates were merge in November 2020. However, I really like this package, so I will accept pull requests over on my fork. If Matteo comes back, I will happily help to merge my changes upstream.

https://github.com/carterbox/torch-radon

I also publish pre-compiled releases of my fork for the conda package manager on the conda-forge channel.

https://anaconda.org/conda-forge/carterbox-torch-radon

NameError: name 'RaysCfg' is not defined

Hi,

The package is installed correctly, but it is not working. There is name 'RaysCfg' is not defined error, how can i fix it?

Thanks!

1. installation

(dlenv) $ python auto_install.py
Checking requirements
Operating System: linux OK
Python version: 3.6 OK
PyTorch: 1.7 OK
CUDA: 10.1 OK
Executing: pip install --force-reinstall https://rosh-public.s3-eu-west-1.amazonaws.com/radon/cuda-10.1/torch-1.7/torch_radon-1.0.0-cp36-cp36m-linux_x86_64.whl
Collecting torch-radon==1.0.0
Using cached https://rosh-public.s3-eu-west-1.amazonaws.com/radon/cuda-10.1/torch-1.7/torch_radon-1.0.0-cp36-cp36m-linux_x86_64.whl (1.0 MB)
Collecting scipy
Using cached scipy-1.5.4-cp36-cp36m-manylinux1_x86_64.whl (25.9 MB)
...
Successfully installed Pillow-8.2.0 alpha-transform-0.0.1 astropy-4.1 cycler-0.10.0 healpy-1.14.0 kiwisolver-1.3.1 matplotlib-3.3.4 numexpr-2.7.3 numpy-1.19.5 pyfftw-0.12.0 pyparsing-2.4.7 python-dateutil-2.8.1 scipy-1.5.4 six-1.15.0 torch-radon-1.0.0 tqdm-4.60.0`

2. when I test 'examples/fbp.py', error shows

NameError Traceback (most recent call last)
in
----> 1 radon = Radon(image_size, angles, clip_to_circle=True)

~/.conda/envs/dlenv/lib/python3.6/site-packages/torch_radon/init.py in init(self, resolution, angles, det_count, det_spacing, clip_to_circle)
167 det_count = resolution
168
--> 169 rays_cfg = RaysCfg(resolution, resolution, det_count, det_spacing, len(angles), clip_to_circle)
170
171 super().init(angles, rays_cfg)

NameError: name 'RaysCfg' is not defined

3. system info

Linux version 3.10.0-862.el7.x86_64 ([email protected]) (gcc version 4.8.5 20150623 (Red Hat 4.8.5-28) (GCC) ) #1 SMP Fri Apr 20 16:44:24 UTC 2018

NVIDIA-SMI 418.152.00
Driver Version: 418.152.00
CUDA Version: 10.1

RuntimeError: CUDA error: invalid device function

NameError: name 'RaysCfg' is not defined with Python 3.10

Hello,

I met this error when using the built version of torch_radon with Google colab, could you please help me to check?

NameError: name 'RaysCfg' is not defined4

Error in Compilation

Output in shell:

(py3) C:\Users\z003zv1a\torch-radon>python setup.py install
Using CUDA_HOME=C:\CUDA\v10.1
�[32mCompiling src\forward.template�[0m
�[34mC:\CUDA\v10.1/bin/nvcc -std=c++11 -ccbin=g++ -Xcompiler -fPIC -Xcompiler -static -Xcompiler -static-libgcc -Xcompiler -static-libstdc++ -I./include -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -DNDEBUG -O3 --generate-line-info --compiler-options -Wall -c src\forward.cu -o objs/cuda/forward.o�[0m
ERROR IN COMPILATION

Operating system: Windows 10
Python version: 3.7

I see that Windows isn't supported oficially. Nevertheless, I'd appreciate any advice to address this issue.

torch_radon_cuda.forward() method returns tensors with all zero

Thank you for your excellent work on this project, but I ran into some problems and wanted to confirm with you.
I have followed your documentation as:

• git clone https://github.com/matteo-ronchetti/torch-radon.git
• cd torch-radon
• python setup.py install

After I run these commands，the terminal displays these commands

and I use python and import torch and import torch_radon_cuda,The program reports no errors.

But,when I use :

  x = torch.randn(2, 1, 512, 512).to("cuda")
  from torch_radon import Radon
  angeles = np.linspace(0, np.pi, 512, endpoint=False)
  radon = Radon(512,angeles)
  pro = radon.forward(x)

I found that the pro here is a tensor with all zero, I also converted this tensor into a PIL image to save, the picture is all black.
I want to know where is the problem.
Hope for your reply

FBP with sinogram obtained by real equipment

I tried to reconstruct sinogram (obtained with real equipment) with torch-radon FBP method. But I only get first image. It may contain a very small object. The second image is obtained by another reconstruction method.

I used "torch_radon.RadonFanbeam()".
There are parameters.

resolution = 512
source_distance = 55 [mm]
det_distance = 148 [mm]
det_spacing = 0.276890625 [mm/pixel]

I wonder if we have to transform parameters when I apply torch-radon FBP for real data.
How can I solve this problem? Thanks.

No module "torch" when run "nosetests tests/ "

I have torch==1.4.0 installed. cuda =10.0 python==3.7 (I installed torch-radon with "python setup.py install").

Example-demo visual_sample.py does not work because the BaseRadon.backprojection does not have a 'extend' argument

Hi

Thank you for publishing this code! It is truly amazing.
After I solved the torch.rfft torch.irfft incompatibility (I want to believe successfully) using two wrapper functions (I post them below) I tried to run the visual_sample.py example but I got the following error

Traceback (most recent call last):
  File "/home/user/torch-radon/examples/visual_sample.py", line 29, in <module>
    backprojection = radon.backprojection(sinogram, extend=True)
  File "/home/user/miniconda3/lib/python3.9/site-packages/torch_radon-1.0.0-py3.9-linux-x86_64.egg/torch_radon/utils.py", line 30, in wrapped
    y = f(self, x, *args, **kwargs)
TypeError: backprojection() got an unexpected keyword argument 'extend'

Should I just delete the 'extend' argument?

The wrapper functions I used to solve the torch.rfft and torch.irfft incompatibility:

def rfft_wrapper(input, signal_ndim, normalized=False):
  if signal_ndim < 1 or signal_ndim > 3:
    print("Signal ndim out of range, was", signal_ndim, "but expected a value between 1 and 3, inclusive")
    return

  dims = (-1)
  if signal_ndim == 2:
    dims = (-2, -1)
  if signal_ndim == 3:
    dims = (-3, -2, -1)

  norm = "backward"
  if normalized:
    norm = "ortho"

  return torch.view_as_real(torch.fft.fftn(input, dim=dims, norm=norm))

def irfft_wrapper(input, signal_ndim, normalized=False):
  if signal_ndim < 1 or signal_ndim > 3:
    print("Signal ndim out of range, was", signal_ndim, "but expected a value between 1 and 3, inclusive")
    return

  dims = (-1)
  if signal_ndim == 2:
    dims = (-2, -1)
  if signal_ndim == 3:
    dims = (-3, -2, -1)

  norm = "backward"
  if normalized:
    norm = "ortho"

  return torch.view_as_real(torch.fft.ifftn(input, dim=dims, norm=norm))

NameError: name 'RaysCfg' is not defined

Hi, I am trying to install the torch-radon library and came into this problem, although the installation process finished and didn't report any error. I follow the author's instructions in a closed issue and have some trials but didn't fix it.

1. I got the error "NameError: name 'RaysCfg' is not defined" when running Radon.
2. I follow the instruction as above and check "from torch_radon_cuda import RayCfg". The same issue came out: "ImportError: /lib/x86_64-linux-gnu/libm.so.6: version GLIBC_2.27 not found".
3. I check the "ldd --version". It is "ldd (Ubuntu GLIBC 2.23-0ubuntu10) 2.23" which <2.27 as suggested.
4. I also try to install the new version by running "wget -qO- https://raw.githubusercontent.com/matteo-ronchetti/torch-radon/v2/auto_install.py | python - " as suggested, but the same errors aforementioned still remain.
5. I also try to compile the library from source. But the compilation is failed and the reported error is "nvcc fatal : Unsupported gpu architecture 'compute_70'".

I very much appreciate any help and suggestion that how this library could be installed and run successfully. Thanks!

How to conduct FBP with fan beam geometry？

I did projection, filtering, and backprojection to a 2D 512×512 image but the reconstruction image is far more different from the original image. Here is the main codes:

 radon = RadonFanbeam(512, angles, 512, 512, det_count=739, det_spacing=1.0, clip_to_circle=False)
 img = torch.FloatTensor(img).to(torch.device('cuda')).unsqueeze(0).unsqueeze(0)
 our_fp = radon.forward(img)
 our_bp = radon.backprojection(radon.filter_sinogram(our_fp))

What's the right way to conduct FBP with fan beam geometry？Thanks a lot.

backprop issues after forward projection

Hi,

First of all thanks for the great library. It's very intuitive to use 👍

However, I have an issue with backward() in torch not working after the forward projection, and some permutations to the result, due to the contiguousy issue.

The task that I'm trying to do here deals with 3D objects, and I guess I can also do that since torch-radon supports batch dimensions. If I have a 3D object of shape (batch, channel, h, w, d), where h refers to height, w refers to width, and d refers to depth, I have to permute the dimensions before computing forward projection, and after the projection, the sinogram (or 2D projections in this case) will have the shape (d, N, det) where N is the number of projection angles, det is the number of detector pixels. For me to see this as multiple 2D projections, I need to permute the dimensions again so that it has the shape (batch, N, det, d). I have written a code which produces the same error I am encountering.

Code

    device = 'cuda:0'

    img_size = 128
    n_angles = 3
    geometry = 'parallel'

    # Parallel beam geometry
    if geometry == 'parallel':
        angles = np.linspace(0, np.pi, n_angles, endpoint=False)
        radon = Radon(img_size, angles, clip_to_circle=False)

    l1_loss = nn.L1Loss(reduction='mean')
    #########################################################
    # Case 1: No ERROR
    #########################################################
    # Random 3D object in pytorch dimensions - (b x c x h x w x d)
    obj = torch.randn([1, 1, 128, 128, 128], requires_grad=True).to(device)

    target = torch.ones([1, 3, 128, 128]).to(device)
    target = proj_bpxy_to_ypx(target)

    sino = radon.forward(vol_bchwz_to_zhw(obj))
    loss = l1_loss(sino, target)
    loss.backward()

    #########################################################
    # Case 2: No ERROR
    #########################################################
    # Random 3D object in pytorch dimensions - (b x c x h x w x d)
    obj = torch.randn([1, 1, 128, 128, 128], requires_grad=True).to(device)

    target = torch.ones([1, 3, 128, 128]).to(device)

    sino = proj_ypx_to_bpxy(radon.forward(vol_bchwz_to_zhw(obj)))
    loss = l1_loss(sino, target)
    loss.backward()

the helper functions are defined as below

def proj_bpxy_to_ypx(tensor):
    """
    Ex. 3-view
    From ~ Input projection     : [3(p), 128(x), 128(y)] or [1(b), 3(p), 128(x), 128(y)]
    To   ~ TorchRadon projction : [128(y), 3(p), 128(x)]
    """
    assert isinstance(tensor, torch.Tensor), 'input must be type torch tensor'
    if tensor.dim() == 4:
        tensor = tensor.squeeze()
    return tensor.permute(2, 0, 1).contiguous()


def proj_ypx_to_bpxy(tensor):
    """
    Ex. 3-view
    TorchRadon projction : [128(y), 3(p), 128(x)]
    Input projection     : [1(b), 3(p), 128(x), 128(y)]
    """
    assert isinstance(tensor, torch.Tensor), 'input must be type torch tensor'
    assert tensor.dim() == 3, f'Dimension must be e.g.[128(x), 3(p), 128(y)]. Received {tensor.shape}'
    y, p, x = tensor.shape
    return tensor.permute(1, 2, 0).contiguous().view(1, p, x, y).contiguous()


def vol_zhw_to_bchwz(tensor):
    """
    Ex.
    TorchRadon BP : [z, h, w]
    Volume shape for 3D Unet Generator: [b, c, h, w, z]
    """
    assert isinstance(tensor, torch.Tensor), 'input must be type torch tensor'
    assert tensor.dim() == 3, f'Dimension must be e.g.[z, h, w]. Received {tensor.shape}'
    z, h, w = tensor.shape
    return tensor.permute(1, 2, 0).contiguous().view(1, 1, h, w, z)


def vol_bchwz_to_zhw(tensor):
    """
    Ex.
    Volume shape for 3D Unet Generator: [b, c, h, w, z]
    TorchRadon BP : [z, h, w]
    """
    assert isinstance(tensor, torch.Tensor), 'input must be type torch tensor'
    assert tensor.dim() == 5, f'Dimension must be e.g.[b, c, h, w, z]. Received {tensor.shape}'
    b, c, h, w, z = tensor.shape
    return tensor.squeeze().contiguous().view(z, h, w)

For the Case 1, there is no error. However, if I run Case 2, I get the following error message.

RuntimeError: x must be contiguous

I guess you usually solve this error with contiguous(), https://stackoverflow.com/questions/48915810/pytorch-contiguous and I've tried putting this to any possible places before and after transpose operations, but it still produce errors. Am I doing something wrong? Could you help me fix this issue?

Thanks a lot in advance

compile error

Compiling src/symbolic.cpp
g++ -std=gnu++11 -fPIC -D_GLIBCXX_USE_CXX11_ABI=0 -Wall -lcuda -lcudart -I./include -DNDEBUG -O3 -c src/symbolic.cpp -o objs/symbolic.o
In file included from ./include/symbolic.h:5,
from src/symbolic.cpp:1:
./include/parameter_classes.h:98:5: error: ‘dim3’ does not name a type
98 | dim3 get_block_dim() const;
| ^~~~
./include/parameter_classes.h:100:5: error: ‘dim3’ does not name a type
100 | dim3 get_grid_size(int x, int y = 1, int z = 1) const;

Conebeam operator

Hi, @matteo-ronchetti ,

Is the conebeam operator a 3D radon transform?

best wishes,
Thermal

Support for higher version of Pytorch/CUDA

Hi.

I find torch-radon library extremely helpful in accelerating forward/backward projection.

It worked well on machines with CUDA 10.2. However, when I tried to run codes on machines with CUDA 11.x, I got the following error message：

Checking requirements
Operating System: linux OK
Python version: 3.7 OK
PyTorch: 1.1 ERROR
Precompiled packages are build for PyTorch 1.5 to 1.7
Consider manually compiling torch-radon

It turns out that this conflicts with the CUDA requirements for pytorch (1.11+), and I am wondering if there's a plan for an updated version of torch-radon that could solve this problem?

Thanks!

Mimimum version of cudatoolkit and pytorch

What are the minimum possible versions of pytorch and cudatoolkit when building from source.

I receive a lot of errors when building using Pytorch 1.1 and cudatoolkit 9.0, which I don't have for Pytorch 1.7 and cudatoolkit 10.

Due to hardware restrictions it is not possible for me to install newer versions. Does it mean that it is hopeless?

Importing exception

I installed torch_radon successfully through:

wget -qO- https://raw.githubusercontent.com/matteo-ronchetti/torch-radon/master/auto_install.py  | python -

and wrote the below codes for testing:

import numpy as np
from torch_radon import Radon, RadonFanbeam
my_radon = Radon(362, np.linspace(0, np.pi, 60, endpoint=False))

But I got:

Importing exception
Importing exception
Traceback (most recent call last):
  File "model.py", line 3, in <module>
    my_radon = Radon(362, np.linspace(0, np.pi, 60, endpoint=False))
  File "/usr/local/lib/python3.6/dist-packages/torch_radon/__init__.py", line 169, in __init__
    rays_cfg = RaysCfg(resolution, resolution, det_count, det_spacing, len(angles), clip_to_circle)
NameError: name 'RaysCfg' is not defined
Exception ignored in: <bound method BaseRadon.__del__ of <torch_radon.Radon object at 0x7f12d9f51e80>>
Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/torch_radon/__init__.py", line 140, in __del__
    self.noise_generator.free()
AttributeError: 'Radon' object has no attribute 'noise_generator'

What should I do?

About the fbp

why in torch-radon/examples/visual_sample.py , the fbp result is radon.backprojection(filtered_sinogram) * np.pi / n_angles,
but in torch-radon/examples/visual_sample.py, the fbp result is radon.backprojection(filtered_sinogram)?
you give two different handle ways

RuntimeError: CUDA error: no kernel image is available for execution on the device

CUDA 11.0，Torch1.7.1

The CUDA error: no kernel image is available for execution on the device was reported when I run the examples. The same example run successfully on my PC, but the above error was reported on the workstation.

Import Exception (missing torch_radon_cuda.py?)

Hi there,

I encountered an error saying " Import Exception" after install it via "wget -qO- https://raw.githubusercontent.com/matteo-ronchetti/torch-radon/master/auto_install.py | python -". The main problem seems to be the missing import of "torch_radon_cuda".

There is a code in torch_radon/init.py

try:
import torch_radon_cuda
from torch_radon_cuda import RaysCfg
except Exception as e:
print("Importing exception")

However, the file "torch_radon_cuda" is completely missing.

Could you have a look at this?
Thanks!

Best,
Di

CUDA error at cudaMemcpy3D(&myparms)

I was using torch-radon v2 branch (which is installed using: wget -qO- https://raw.githubusercontent.com/matteo-ronchetti/torch-radon/v2/auto_install.py | python - provided in Issue #23).

But when I try to run some test code I got the following error:

CUDA error at cudaMemcpy3D(&myparms) (src/texture.cu:167) error code: 1, error string: invalid argument

The GPU was not occupied (so it shouldn't be a problem of OOM?), and my test code is like the following:

import torch
import numpy as np
from torch_radon.radon import FanBeam


if __name__ == '__main__':
    import os
    import numpy as np
    img = np.random.randn(256, 256)
    
    os.environ['CUDA_VISIBLE_DEVICES'] = '4'
    print(torch.cuda.device_count())
    print(torch.cuda.is_available())
    angles = np.linspace(0, 2 * np.pi, 360)
    img = torch.from_numpy(img).to("cuda").float().reshape(1, 1, 256, 256)
    print(img.shape, img.min(), img.max())
    
    tool = FanBeam(det_count=768, angles=angles, src_dist=1075)
    print('FanBeam tool initialized')
    print(tool.forward(img, angles))
    

The full error messages are:

1
True
torch.Size([1, 1, 256, 256]) tensor(-4.5295, device='cuda:0') tensor(4.8204, device='cuda:0')
FanBeam tool initialized
CUDA error at cudaMemcpy3D(&myparms) (src/texture.cu:167) error code: 1, error string: invalid argument

Any advice on solving this issue? Thanks!

ImportError: `GLIBC_2.27' not found

Hi, after installing torch-radon successfully, I encountered a problem,
"
File "/usr/local/lib64/python3.6/site-packages/torch_radon/init.py", line 6, in
import torch_radon_cuda
ImportError: /usr/lib64/libm.so.6: version `GLIBC_2.27' not found (required by /usr/local/lib64/python3.6/site-packages/torch_radon_cuda.cpython-36m-x86_64-linux-gnu.so)
"
My env is -python3.6-cuda10.1-cudnn7.6-pytorch1.5.

I wonder if this problem is caused since the origin .whl is compiled on a higher-vision of Linux.

Could you help me fix this problem？ thanks!

ImportError: cannot import name 'setup'

Hi Matteo,

Nice work!
I encountered some error messages below when I tried to install your package from source. Do you know how to fix this?

Thanks,
Dongdong

---Steps:
$ git clone https://github.com/matteo-ronchetti/torch-radon.git
$ cd torch-radon
$ python setup.py install
---Errors:
Traceback (most recent call last):
File "setup.py", line 1, in
from setuptools import setup
ImportError: cannot import name 'setup'

reconstruction performance

hello, when testing the torch-radon algorithm, I found a performance-related problem.

if I use radon()->radon.forward()->radon.filtered()->radon.backprojection(), the reconstructed image has a low PSNR performance (i.e. big MSE) compared to offical matlab code (the same setting).

For example, for a CT image, the reconstruction performance of torch-radon is 29.39dB, while that of matlab is nearly 40 dB.
Obviously, this reconstruction is not so satisfactory due to low PSNR.

So, could you tell me how to boost the basic performance of torch-radon? (without adding cnn modules and training?)
My goal is to obtain high-quality sinograms (projections).

Thanks! Appreciate for your help!