https://github.com/zaccharieramzi/understanding-unets

Learnlets repository
https://github.com/zaccharieramzi/understanding-unets

convolutional-neural-networks denoising learnlets u-net wavelets

Last synced: about 1 year ago
JSON representation

Learnlets repository

• Host: GitHub
• URL: https://github.com/zaccharieramzi/understanding-unets
• Owner: zaccharieramzi
• Created: 2019-04-01T14:28:13.000Z (about 7 years ago)
• Default Branch: master
• Last Pushed: 2023-03-25T00:35:56.000Z (about 3 years ago)
• Last Synced: 2025-03-30T09:32:10.510Z (about 1 year ago)
• Topics: convolutional-neural-networks, denoising, learnlets, u-net, wavelets
• Language: Jupyter Notebook
• Homepage:
• Size: 79.8 MB
• Stars: 9
• Watchers: 1
• Forks: 4
• Open Issues: 7
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Learnlets

[![Build Status](https://travis-ci.com/zaccharieramzi/understanding-unets.svg?token=wHL4tmyGD3TP6bSo6Mdh&branch=master)](https://travis-ci.com/zaccharieramzi/understanding-unets)

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 |