wdika/mridc

Data Consistency Toolbox for Magnetic Resonance Imaging

mridc's Issues

Add image logging for wandb

Is your feature request related to a problem? Please describe.
Go to the base reconstruction class and the logging func.

Describe the solution you'd like
It would be nice to log images in wandb as in tensorboard atm.

[FR] Use PyTorch Lightning

Is your feature request related to a problem? Please describe.
Training and logging should be more concise and unified for every model.

Describe the solution you'd like
Use PyTorch Lightning.

remove fixed data imports specific for the paper

Merge RSS for coil combination among inputs and targets

Is your feature request related to a problem? Please describe.
For example, the option for the inputs can be defined here.
We need to make sure this applies here and across the repo.

Describe the solution you'd like
Use the coil combination func everywhere.

Define coil and spatial dimensions, instead of dataset type.

Is your feature request related to a problem? Please describe.
In the config files, the dataset_type argument should be removed (for example here).

Describe the solution you'd like
Instead, we should define the coil and spatial dimensions.

[BUG] Batch size > 1 is not allowed

Describe the bug
--batch_size > 1 is not allowed. It also mixes with the input's slice dimension.

To Reproduce
Train a cirim or e2evn or unet with --batch_size > 1.

Expected behavior
--batch_size > 1 should be allowed. When training on a dataset with varying matrix sizes though it should be limited to 1.

Environment
Operating System: ubuntu-latest Python Version: >= 3.9 PyTorch Version: >= 1.9

Coil dim in the llg of the RIM is fixed

Describe the bug
The coil dimension in the log-likelihoog gradient computation is fixed to 1.

To Reproduce
Steps to reproduce the behavior:
Go to rim_utils

Expected behavior
This should be dynamically set by the argument of the function.

[FR] Use weights and biases

Is your feature request related to a problem? Please describe.
Hyperparameters search options need. Logging needs to be refined.

Describe the solution you'd like
Use weights and biases.

[FR] Generalize or constrain the ismrmrd header metadata for every dataset

Is your feature request related to a problem? Please describe.
The fastMRI dataset contains metadata stored in ismrmrd header format. This is true only for some datasets and not always in general.

Describe the solution you'd like
This function _retrieve_metadata should be generalized or constrained.

onnxruntime 1.11.1 doesn't support python 3.10

Is your feature request related to a problem? Please describe.
As title says onnxruntime 1.11.1 doesn't support python 3.10.

Describe the solution you'd like
Upgrade when onnxrutime updates.

Generalize the FFT type

Is your feature request related to a problem? Please describe.
The FFT type (for example here and here and here) can be defined multiple times and it is not always clear how normalization works.

Describe the solution you'd like
The fft_normalization and fft_dim should be leveraged better, to distinguish between several different options.

Add improvements [new models] to README

Fix/allow batch_size > 1

Describe the bug
A batch size of 1 is only supported at the moment. This behavior is not indeed and should be changed.

To Reproduce

mridc/mridc/collections/reconstruction/models/base.py

Line 287 in 32a3e18

batch_size=1,

Expected behavior
batch_size>1 should be allowed.

Desktop (please complete the following information):

• OS: Linux

Additional context
Add any other context about the problem here.

add documentation for quantitative module

merge to main & upgrade to 0.2.0

[FR] Create global arguments file and inherit them

Is your feature request related to a problem? Please describe.
Every train and run script duplicates global arguments.

Describe the solution you'd like
Global arguments should be on a separate file and be inherited when required.

Describe alternatives you've considered
This might be also addressed through PL and WANDB.

[FR] Unify Fast Fourier Transform across the repo

Is your feature request related to a problem? Please describe.
Fast Fourier Transform type, dim, and normalization should be consistent across the repo.

Describe the solution you'd like
Define fft_type, fft_dim, and fft_normalization across the repo.

• fft_type options should be "orthogonal" and "backward".
• fft_normalization options should be "orthogonal", "backward", and None.
• fft_dim should always be the last two dimensions or the two last before the complex dimension [..., 2], when this exists.

The center_fractions argument can be confusing for non cartesian masking

Describe the bug
Not precisely a bug, but this argument should be generalized among different masking functions.

To Reproduce
Steps to reproduce the behavior:

1. Go to MaskFunc.
2. See how it acts for random cartesian masking and how for gaussian masking.

Expected behavior
For example, in gaussian masking, this argument acts as FWHM. This should be clearly formalized.

re test the qcirim implementation to work without absolute inputs

[BUG] Check if half_scan_percentage alters the chosen acceleration factors / Add option for 1D masks

Describe the bug
half_scan_percentage works only with 2D masks.

To Reproduce
Duplicate a test function for 1D masking like this and add half_scan_percentage > 0.

Expected behavior
It should be working on both 1D and 2D masks. Add this option for 1D masking as well.

Environment
Operating System: ubuntu-latest Python Version: >= 3.9 PyTorch Version: >= 1.9

[FR] Parse arguments for train and run from yaml file

Describe the solution you'd like
Use yaml files to parse arguments for training and running methods.

hydra-core==1.2.0 is not supported

Describe the bug
hydra-core==1.2.0 is not supported.

To Reproduce
Steps to reproduce the behavior:
Try to run train/inference with mridc.launch for any model.

Expected behavior
mridc.launch should not return any error for _run_hydra