Learnlets

Learnlets are a way to learn a filter bank rather than design one like in the curvelets.

This filter bank will be learned in a denoising setting with backpropagation and gradient descent.

Requirements

The requirements are listed in learning_wavelets/requirements.txt.

Use

The learnlets are defined in learning_wavelets/learnlet_model.py, via the class Learnlet.

You can use different types of thresholding listed in learning_wavelets/keras_utils/thresholding.py.

List of saved networks

Exact reconstruction notebook

Model id	Params
learnlet_dynamic_st_bsd500_0_55_1580806694	the big classical network, with 256 filters + identity
learnlet_subclassing_st_bsd500_0_55_1582195807	64 filters, subclassed API, exact recon forced

No threshold notebook

Model id	Params
learnlet_dynamic_st_bsd500_0_55_1580806694	the big classical network, with 256 filters + identity

Different training noise standard deviations notebook

Model id	Params
learnlet_dynamic_st_bsd500_0_55_1580806694	the big classical network, with 256 filters + identity
learnlet_dynamic_st_bsd500_20_40_1580492805	same with training on 20;40 noise std
learnlet_dynamic_st_bsd500_30_1580668579	same with training on 30 noise std
unet_dynamic_st_bsd500_0_55_1576668365	big classical unet with 64 base filters and batch norm
unet_dynamic_st_bsd500_20.0_40.0_1581002329	same with training on 20;40 noise std
unet_dynamic_st_bsd500_30.0_30.0_1581002329	same with training on 30 noise std

General comparison

Model id	Params
learnlet_dynamic_st_bsd500_0_55_1580806694	the big classical network, with 256 filters + identity
unet_dynamic_st_bsd500_0_55_1576668365	big classical unet with 64 base filters and batch norm