openspyrit/spyrit

A Python toolbox for deep image reconstruction, with emphasis on single-pixel imaging.

spyrit's Issues

Patterns for CompNet initialization

With Walsh being the new method for generating the patterns, the following line should probably be changed to create the matrix H using wh.walsh2_matrix when no matrix is given during initialization of a CompNet:

spyrit/spyrit/learning/model_Had_DCAN.py

Line 561 in 0ff170f

H = Hadamard_Transform_Matrix(self.n)

Remove H_adjoint Linear attribute

spyrit/spyrit/core/meas.py

Line 82 in a4c92d1

# adjoint (Remove?)

Remove fht

In #29

@nducros

Great. So this is the right time to discard it!

You should only need to replace this line with
H = wh.walsh2_matrix(n)/n

Both Stat_h and optim_had are no longer necessary as I reimplemented the walsh version of both functions (already included)

I started a new branch here
#34

@romainphan Should we replace the definition of meas2img by that of meas2img2, which seems more general?

          @romainphan Should we replace the definition of meas2img by that of meas2img2, which seems more general?

spyrit/spyrit/misc/sampling.py

Line 47 in 607d027

def meas2img2(meas: np.ndarray, Mat: np.ndarray) -> np.ndarray:

Originally posted by @nducros in #133 (comment)

generalised_orthogonal_debug.ipynb

Document the spyrit functions of generalised_orthogonal_debug.ipynb

Package spyrit

I open this issue to put together all information to create the package for spyrit.

• Add FHT (https://github.com/nbarbey/fht) as submodule. The pypi package is not available for all python version and distribution. See (https://pypi.org/project/fht/#files). It seems that it works for Linux and python 3.6 only.
• Compile and create the wheel localy
• Create a github actions to create the wheel on ManyLinux, Mac and (Windows ?) and multiple python versions
• Push on pypi the github actions wheel

Acqusition_debug

Document the spyrit functions of https://github.com/openspyrit/spyrit-examples/blob/master/Debug_version_2/Acqusition_debug.ipynb

Measurement simulation (in meas.py)

Instead of having LinearDynamic inheriting of Linear, what about having Linearinheriting of LinearDynamic ? To me this seems logical because looking at a still picture is a particular case of a moving picture.
Also, the adjoint / pinv methods that we currently inherit from Linear would be redefined only in the child class, which would avoid having useless methods.
@nducros ?

Replace setup.py by pyproject.toml

pyproject.toml

https://pip.pypa.io/en/stable/reference/build-system/pyproject-toml/

Error when creating a DenoiCompNet

Following the updates from PR #37, where misc/statistics was added, I am now getting an error when creating a DenoiCompNet using spas. It seems Cov2Var cannot be found.

Traceback (most recent call last):

  File "C:\Users\beneti\spas\scripts\acquisition-recon-sequential.py", line 63, in <module>
    model, device = setup_reconstruction(cov_path, mean_path, H, model_root, network_params)

  File "c:\users\beneti\spas\spas\reconstruction_nn.py", line 146, in setup_reconstruction
    model = DenoiCompNet(

  File "c:\users\beneti\spyrit\spyrit\learning\model_Had_DCAN.py", line 598, in __init__
    super().__init__(n, M, Mean, Cov, variant, N0, sig, H, Ord)

  File "c:\users\beneti\spyrit\spyrit\learning\model_Had_DCAN.py", line 531, in __init__
    super().__init__(n, M, Mean, Cov, variant, H, Ord)

  File "c:\users\beneti\spyrit\spyrit\learning\model_Had_DCAN.py", line 336, in __init__
    Ord = Cov2Var(Cov)

NameError: name 'Cov2Var' is not defined

OpenCV

Remove OpenCV dependency and change to PIL
But some functions cannot be replaced: https://github.com/openspyrit/spyrit/blob/master/spyrit/misc/disp.py#L308

-> Change everything in the test at firsdt

Update readme in spyrit (admin user) ?

Create a branch for testing the documentation update.

Method `forward_Pinv` is not defined in compNet

The method forward_Pinv() is not defined in compNet, but is called by evaluate_Pinv(). As seen in:

spyrit/spyrit/learning/model_Had_DCAN.py

Line 742 in 0ff170f

x = self.forward_Pinv(x, b, c, h, w)

Besides that, is evaluate_Pinv() supposed to call forward_acquire() before calling forward_Pinv()?

Pinv_net_debug.ipynb

Document the spyrit functions of Pinv_net_debug.ipynb

Test

Reconstruction with multiple channels

In spyrit.learning.model_Had_DCAN, the function forward_N0_maptoimage_expe works for reconstruction of a batch (b != 0) but does not work for a multichannel reconstruction (c != 0).
The problem comes from the following lines:

N0_est = self.max(x_est);
N0_est = N0_est.view(b*c,1,1);
N0_est = N0_est.repeat(1,1,self.M)
x = torch.div(x,N0_est)

N0_est has the dimensions (b*c, 1, self.M) , whereas x has the dimensions (b, c, self.M).

It seems a possible solution would be to convert x to the same dimensions as N0_est at the beginning of the function.

simplify init of compNet

Add default parameters
No need for Cov and Mean for the evaluation of trained networks

Simplify compNet (no need to: (i) rescale between [-1 1] and (ii) split)

builds are back

P... Je me suis fais avoir sur towards_v2 ! Je pensais être sur "ma" branche et aussi que towards_v2 était protégée, mais non.
Je suis vraiment confus Thomas, j'aurais dû vérifier je connais la procédure maintenant

@TALAOUI : Ne pull pas surtout !

Import errors

I am getting some errors due to import issues.

• In spas/misc/statistics.py the following line is used import spyrit.misc.walsh_hadamard as wh but then some function calls do not start by wh as in:

  spyrit/spyrit/misc/statistics.py

  Line 183 in 8c80206

  H = walsh_matrix(len(img))

  spyrit/spyrit/misc/statistics.py

  Line 123 in 8c80206

  h_img = walsh_ordered2(img, H1d);

  In this case it seems I cannot train a new network because Stat_had() will fail.

• As mentioned in #39 some imports in spas/learning/model_Had_DCAN.py do not work, I have noticed 3 factors that may contribute to this: circular imports, imports using * (star imports) and relative and absolute imports being used in a same file.

• In a same file, imports like from ..misc.pattern_choice import Hadamard, matrix2conv (which is a relative import) and from spyrit.misc.statistics import * (which is not) are used, which I believe is not a best practice.

I believe it would be a good idea to review most of the imports in all relevant files, choose to use a method, i.e. relative imports, avoid star imports as much as possible and remove unnecessary legacy imports in file headers.

Clean up

Following the surpression of 'restructured', also delete ancestors such as:
model_DCAN.py
model_Had_DCAN.py
model_Had_DCAN_1.py

Renaming of walsh_ordered

Hello,
It seems that some methods related to Hadamard have been renamed, causing the deprecation of some code.

For example, here
and there

walsh_ordered is called but I could not find a definition of walsh_ordered.

Probably walsh_ordered should be replaced by its new name?

Spyrit automatic documentation

Should we have separate html pages for each module ? The core package documentation page is way too long and could be divided into subpages for each package. Same for the misc package.
What do you think ?
@tbaudier @nducros

Provide minimal working examples

Explain in readme and/or ddocumentation

• How to allocate compNet?
• How to restore a compNet?
• How to train a compNet?

pip install spyrit > no tutorials

with user installation of spyrit (pip install spyrit), the tutorials and images don't come along
@tbaudier

No more images/outputs in documentation

E.g., see here
@romainphan: Is it your last commit?

Merge meas2img and meas2img2 into new, more general function

meas2img returns the measurement image for a single measurement vector, while meas2img2 does the same for a batch. Is there a way to provide one unique function that does the same thing ?

Forward_operator_debug

Document the spyrit functions of Forward_operator_debug.ipynb

Merge doc branch

Create welcome page for the spyrit organization

push under windows and fht

git status outputs

modified: fht
modified: spyrit/fht (untracked content)

What should I do?

Can I git add -A and push ? Will it break fht (e.g. under linux)?

Train networks using input with size [b,c,w,h]

Speed up computation of mean and covariance

Using torch tensor as in DLMIS

Simplify noiCompNet: compute covariance directly (no rescaling in [0, N0]) followed by inverse transformations

Network attributes loading

Specify in documentation that load_net does not load all network attributes (check and explain why)

spyrit/spyrit/learning/nets.py

Line 445 in ba3d709

def load_net(title, model, device):

train_Pinv_Net_Unet.py

Document the spyrit functions of train_Pinv_Net_Unet.py

Pinv is nn.Linear while P is a nn.Cond2D

All of them should be the same type (the faster matrix-vector product implementation)

drunet

To get the tuto on drunet, check this branch after the paper branch is merged with master.

pip install spyrit fails on MacOS (High Sierra)

I followed Thomas's recommendation : no more using the cache.
But it is still downloading the .tar.gz and failing.

Is it because there is no wheel for MacOS (High Sierra)?

pip install spyrit fails on CREATIS cluster (Fedora)

I followed Thomas's recommendation : no more using the cache.
But it is still downloading the .tar.gz and failing.

Is it because there is no wheel for Fedora?

Add test 2021_OpticsExpress + update 2020_ISBI in CI

version 0.14 not present on pypi

@nducros the pypi version was not updated with version 0.14
https://pypi.org/project/spyrit/

The version has to be changed here:
https://github.com/openspyrit/spyrit/blob/master/setup.py#L45

Use append to build iterate module in deepEM network

code in the master_Antonio branch

see https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html

Relative import

You should not import relative past:

https://github.com/openspyrit/spyrit/blob/doc/spyrit/acquisiton/load_data.py#L21-L23

API examples

1. Copy the long examples of the API documentation in files (e.g., 'test/test_forward.py')
2. Keep API examples as short as possible (limit the number of variables to a minimum)

Implement the fast walsh hadamard transform in torch

Algorithm https://en.wikipedia.org/wiki/Fast_Walsh%E2%80%93Hadamard_transform

Rename models

DCAN stands for deep convolutional auto-encoder
network, named after Higham et al 2018 paper.
Our models are more general