For Calibratinless reconstruction, we would have to carry out wavelet transform across all the channels.
In order to speed things up, I set up the codes for running linear_op in parallel.
However, this caused the following issue:
In the definition of WaveletN, we define just one linear operator:
In doing so, we are eventually using the same wavelet operator for all the channels in parallel which causes severe races.
Due to this the overall image is bad and has artifacts like phase wrapping in magnitude image.
I will get a PR fixing this, but for now, users are requested to not use n_jobs
Also, this seems like a hole in testing, hence adding a testing tag and I will add a test for the same.
Thanks to @Daval-G in helping in identifying this major issue...
For Calibratinless reconstruction, we would have to carry out wavelet transform across all the channels. In order to speed things up, I set up the codes for running linear_op in parallel. However, this caused the following issue:
In the definition of WaveletN, we define just one linear operator:
https://github.com/CEA-COSMIC/pysap-mri/blob/db04fe2956adaf0b1e84bc58dc99e5f63664286e/mri/operators/linear/wavelet.py#L61-L62
Then we use this in parallel to obtain our images in
_op()function:https://github.com/CEA-COSMIC/pysap-mri/blob/db04fe2956adaf0b1e84bc58dc99e5f63664286e/mri/operators/linear/wavelet.py#L71-L77
In doing so, we are eventually using the same wavelet operator for all the channels in parallel which causes severe races. Due to this the overall image is bad and has artifacts like phase wrapping in magnitude image.
I will get a PR fixing this, but for now, users are requested to not use
n_jobsAlso, this seems like a hole in testing, hence adding a testing tag and I will add a test for the same.Thanks to @Daval-G in helping in identifying this major issue...