About the SART-TV algorithm

ffbfbfdf commented 1 month ago

Hello， author！ I'm having the following issue with Tigre：

Specifications

MATLAB/python version: In this algorithm, the TV constraint is carried out after the SART iteration is completed. I didn't do a sart reconstruction for this, instead I entered a picture recBone, followed by "Bone=im3DDenoise(recBone,'TV',TViter,TVlambda,'gpuids',gpuids); I've found out that Bone is almost all "NAN", how can I avoid this?

ffbfbfdf commented 1 month ago

A simple example is shown in the image.

AnderBiguri commented 1 month ago

Humm strange. What if you make A=rand(5,5,5)?

ffbfbfdf commented 1 month ago

Same as two-dimensional

AnderBiguri commented 1 month ago

Very strange! Can you tell my the cuda version and gpu you are running it on? I am away but I will try to fix this next week.

ffbfbfdf commented 1 month ago

GPU：NVIDIA Geforce RTX 3060 LapTop Gpu

ffbfbfdf commented 1 month ago

In other words, can other versions of CUDA and GPU run normally?

AnderBiguri commented 1 month ago

@ffbfbfdf no idea, the error is strange, so maybe this information helps me debug. I'll try to debug next week, apologies fornt delay

tsadakane commented 1 month ago

Hi, @ffbfbfdf , @AnderBiguri In my environment of Python, I could not reproduce.

import numpy as np
import numpy.random
import tigre
import tigre.algorithms as algs
import tigre.utilities.gpu as gpu
import tigre.utilities.im_3d_denoise

listGpuNames = gpu.getGpuNames()
if len(listGpuNames) == 0:
    print("Error: No gpu found")
else:
    for id in range(len(listGpuNames)):
        print("{}: {}".format(id, listGpuNames[id]))

gpuids = gpu.getGpuIds(listGpuNames[0])
print(gpuids)

A=np.random.randn(1, 5, 5).astype(np.float32)
A[A<0]=0
print(f"A = \n{A}")
B=tigre.utilities.im_3d_denoise.im3ddenoise(A, iter=2, gpuids=gpuids)
print(f"B = \n{B}")

Output

0: NVIDIA GeForce GTX 1060 6GB
{'name': 'NVIDIA GeForce GTX 1060 6GB', 'devices': [0]}
A =
[[[0.         0.         0.         0.45640853 0.8639155 ]
  [1.1695538  0.         1.1146449  0.         1.1266571 ]
  [1.1079429  0.         0.60824865 0.         0.        ]
  [0.8022037  1.425755   1.6020744  0.         0.        ]
  [0.7927743  0.9320209  0.4525195  0.8293371  0.        ]]]
B = 
[[[-0.01066     0.         -0.0143195   0.46933416  0.88098335]
  [ 1.2127553  -0.02081952  1.170058   -0.02458851  1.170128  ]
  [ 1.1404634  -0.02863753  0.6167504  -0.00554393 -0.01026901]
  [ 0.80844307  1.473317    1.6670241  -0.0221613   0.        ]
  [ 0.80587083  0.95015025  0.44248584  0.86300796 -0.00755906]]]

AnderBiguri commented 1 month ago

@tsadakane thanks for testing. Indeed, I suspect this may be perhaps from some behaviour of the code that is CUDA version specific, as I can not reproduce either. Which CUDA are you using?

ffbfbfdf commented 1 month ago

嗨，，在我的 Python 环境中，我无法重现。

import numpy as np
import numpy.random
import tigre
import tigre.algorithms as algs
import tigre.utilities.gpu as gpu
import tigre.utilities.im_3d_denoise

listGpuNames = gpu.getGpuNames()
if len(listGpuNames) == 0:
    print("Error: No gpu found")
else:
    for id in range(len(listGpuNames)):
        print("{}: {}".format(id, listGpuNames[id]))

gpuids = gpu.getGpuIds(listGpuNames[0])
print(gpuids)

A=np.random.randn(1, 5, 5).astype(np.float32)
A[A<0]=0
print(f"A = \n{A}")
B=tigre.utilities.im_3d_denoise.im3ddenoise(A, iter=2, gpuids=gpuids)
print(f"B = \n{B}")

输出

0: NVIDIA GeForce GTX 1060 6GB
{'name': 'NVIDIA GeForce GTX 1060 6GB', 'devices': [0]}
A =
[[[0.         0.         0.         0.45640853 0.8639155 ]
  [1.1695538  0.         1.1146449  0.         1.1266571 ]
  [1.1079429  0.         0.60824865 0.         0.        ]
  [0.8022037  1.425755   1.6020744  0.         0.        ]
  [0.7927743  0.9320209  0.4525195  0.8293371  0.        ]]]
B = 
[[[-0.01066     0.         -0.0143195   0.46933416  0.88098335]
  [ 1.2127553  -0.02081952  1.170058   -0.02458851  1.170128  ]
  [ 1.1404634  -0.02863753  0.6167504  -0.00554393 -0.01026901]
  [ 0.80844307  1.473317    1.6670241  -0.0221613   0.        ]
  [ 0.80587083  0.95015025  0.44248584  0.86300796 -0.00755906]]]

Why are there negative numbers in matrix B？

AnderBiguri commented 1 month ago

@ffbfbfdf probably because the parameters of the denoisers are not appropriate for the toy problem set here. However this is not very relevant to the error you are getting, which is that you get NaNs

tsadakane commented 1 month ago

@tsadakane thanks for testing. Indeed, I suspect this may be perhaps from some behaviour of the code that is CUDA version specific, as I can not reproduce either. Which CUDA are you using?

CUDA 12.2 with VS2022

tsadakane commented 1 month ago

@ffbfbfdf , @AnderBiguri,

In my environment of Matlab, I could reproduce.

>> A=randn(5,5)
A =
    1.0347    0.8884    1.4384   -0.1022   -0.0301
    0.7269   -1.1471    0.3252   -0.2414   -0.1649
   -0.3034   -1.0689   -0.7549    0.3192    0.6277
    0.2939   -0.8095    1.3703    0.3129    1.0933
   -0.7873   -2.9443   -1.7115   -0.8649    1.1093
>> A(A<0)=0
A =
    1.0347    0.8884    1.4384         0         0
    0.7269         0    0.3252         0         0
         0         0         0    0.3192    0.6277
    0.2939         0    1.3703    0.3129    1.0933
         0         0         0         0    1.1093
>> [B]=im3DDenoise(A, 'TV')
B =
  5×5 single matrix
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN

CUDA 12.2, VS2022, Matlab 2022a. Hmm, too many two's

ffbfbfdf commented 1 month ago

@ffbfbfdf probably because the parameters of the denoisers are not appropriate for the toy problem set here. However this is not very relevant to the error you are getting, which is that you get NaNs

Yes, expecting help for you to rewrite the bug and hope you give a solution

ffbfbfdf commented 1 month ago

，，

在我的Matlab环境中，我可以复制。

>> A=randn(5,5)
A =
    1.0347    0.8884    1.4384   -0.1022   -0.0301
    0.7269   -1.1471    0.3252   -0.2414   -0.1649
   -0.3034   -1.0689   -0.7549    0.3192    0.6277
    0.2939   -0.8095    1.3703    0.3129    1.0933
   -0.7873   -2.9443   -1.7115   -0.8649    1.1093
>> A(A<0)=0
A =
    1.0347    0.8884    1.4384         0         0
    0.7269         0    0.3252         0         0
         0         0         0    0.3192    0.6277
    0.2939         0    1.3703    0.3129    1.0933
         0         0         0         0    1.1093
>> [B]=im3DDenoise(A, 'TV')
B =
  5×5 single matrix
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN
   NaN   NaN   NaN   NaN   NaN

CUDA 12.2、VS2022、Matlab 2022a。嗯，两个太多了

Does this issue occur with the python version of the code?

tsadakane commented 1 month ago

Is the input 'A' allowed to be a double precision matrix? In python, it was necessary to use 'astype(float)' to run that code.

ffbfbfdf commented 1 month ago

输入“A”是否允许是双精度矩阵？在python中，必须使用“astype（float）”来运行该代码。

Makes sense, I type in a single precision to get the result, but there is a negative value

AnderBiguri commented 1 month ago

@ffbfbfdf the mathematics allow negative values, so there is no issue with that.

So is the issue double vs single? Or cuda 12.2?

AnderBiguri commented 1 month ago

I guess teh bug to fix here is that im3Dnoise should cast/error

CERN / TIGRE

About the SART-TV algorithm #553

Specifications