You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the Condat-Vu reconstruction algorithm (for example in the reconstruct plugin), the options of reweighting "primal" and "dual" lead to some errors:
With "primal", it calls a method not defined in gradient method and crashes.
With "dual", if you use the wrapper to call the transforms, in some specific cases (transforms with nb_bands > 1), it applies wrongly the linear operator and the result of the reconstruction is deteriorated.
I haven't find a fix yet, as reweighting is out of my understanding.
For now, Sylvain has removed it in his PR factory, as it isn't used by anyone.
The text was updated successfully, but these errors were encountered:
bensarthou
changed the title
Problems by using the transforms bindings with reweighting
Problems when using the transforms bindings with reweighting
May 7, 2018
The "dual" reweighting uses linear_op.transform object to obtain reweighting factor for each band, this is not scalable for all cases, especially for undecimated wavelet which I implemented for issue #4 .
Both reweightings are incompatible for multichannel cases. Both would ideally have different reweighting per channel per band.
In the Condat-Vu reconstruction algorithm (for example in the reconstruct plugin), the options of reweighting "primal" and "dual" lead to some errors:
With "primal", it calls a method not defined in gradient method and crashes.
With "dual", if you use the wrapper to call the transforms, in some specific cases (transforms with nb_bands > 1), it applies wrongly the linear operator and the result of the reconstruction is deteriorated.
I haven't find a fix yet, as reweighting is out of my understanding.
For now, Sylvain has removed it in his PR factory, as it isn't used by anyone.
The text was updated successfully, but these errors were encountered: