KeraStroke

KeraStroke is a Python package that implements "Post-Back-propagation Weight Operations", or "PBWOs"; generalization-improvement techniques for Keras models in the form of custom Keras Callbacks. These techniques function similarly but have different philosophies and results. The techniques are:

Stroke: Re-initializaing random weight/bias values.
Pruning: Reducing model size by setting weight/bias values that are close to 0, to 0.
NeuroPlast: Re-initializing any weight/bias values that are 0 or close to 0.

Stroke is modeled after seizures, which send random electrical signals throughout the brain, sometimes causing damage to synapses.

NeuroPlast is modeled after the concept of neuroplasticity, when neurons that no longer have a primary function begin to rewire to improve another function. I started working on NeuroPlast after I read the work done by Blakemore and Cooper on horizontal/vertical line receptor neurons in the brains of cats.

If you'd like to see the tests I'm performing with KeraStroke, you can view my testing repository here.

KeraStroke 2.0.0 marks when I really started putting work into the project. I've made an effort to comment more, clean my code up, and make the package easier to understand overall without sacrificing utility.

Limitations

KeraStroke is still in the development phase. Heavy testing has been done on Dense nets, but little testing has been done on CNNs and no testing has been done on RNNs. As of 2.1.0, CNNs are functioning properly in KeraStroke! The issue with previous versions had to do with the way the callback would retrieve the weights from the models. The callbacks perform significantly better on DenseNets, but could still find use in CNNs. I'm working on this, but will definitely need the help. Please see the github page or contact me to contribute to the project.

Stroke

The goal of the Stroke callback is to re-initialize weights/biases that have begun to contribute to overfitting.

Parameters:

set_value: re-initialized weights will be set to this value, rather than a random one
low_bound: low bound for weight re-initialization
high_bound: high bound for weight re-initialization
volatility_ratio: percentage of weights to be re-initialized
cutoff: number of epochs to perform PBWOs
decay: Every epoch, v_ratio is multiplied by this number. decay can be greater than 1.0, but v_ratio will never exceed 1.0
do_weights: perform stroke on weights
do_biases: perform stroke on biases

Pruning

The goal of the Pruning callback is to nullify weights/biases that are effectively 0.

Parameters:

set_value: The value that pruned weights will be set to
min_value: The lowest value a weight/bias can be to be oeprated on
max_value: The highest value a weight/bias can be to be operated on
cutoff: number of epochs to perform PBWOs
do_weights: perform pruning on weights
do_biases: perform pruning on biases

NeuroPlast

The goal of the NeuroPlast callback is to randomly re-initialize weights/biases that are effectively 0.

Parameters:

set_value: re-initialized weights will be set to this value, rather than a random one
min_value: lowest value a weight/bias can be to be operated on
max_value: highest value a weight/bias can be to be operated on
low_bound: low bound for weight re-initialization
high_bound: high bound for weight re-initialization
cutoff: number of epochs to perform PBWOs
do_weights: perform neuroplast on weights
do_biases: perform neuroplast on biases

Usage

KeraStroke Callbacks can be used like any other custom callback. Here's a basic example:

from kerastroke import Stroke
model.fit(X, y, 
          epochs=32, 
          callbacks=[Stroke()])

License
- OSI Approved :: MIT License
Operating System
- OS Independent
Programming Language
- Python :: 3

Release files for kerastroke 2.1.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for kerastroke 2.1.1
File	Size	Uploaded
kerastroke-2.1.1.tar.gz	120.3 kB	Apr 28, 2020	Details

Built distribution (wheel)

Table of built distributions (wheels) for kerastroke 2.1.1
File	Interpreter	ABI	Platform	Reset
kerastroke-2.1.1-py3-none-any.whl 7.0 kB Apr 28, 2020	Python 3	none	any	Details

Total release size: 127.4 kB

Release files / kerastroke-2.1.1.tar.gz

Download URL	kerastroke-2.1.1.tar.gz
Size	120.3 kB
Tags	Source
SHA-256 checksum How to use checksums	`3eb25f3d698de7ae85b68a143fd62e323d455e29f595b413128822c33ef8620e`
BLAKE2b-256 checksum How to use checksums	`3c4ed74e798142e43eb3ec7129cce6704b7edaaded710e0dea8f1fb79e9d7175`
Upload date	Apr 28, 2020
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/3.1.1 pkginfo/1.5.0.1 requests/2.11.1 setuptools/45.1.0 requests-toolbelt/0.9.1 tqdm/4.31.0 CPython/3.7.7`

Release files / kerastroke-2.1.1-py3-none-any.whl

Download URL	kerastroke-2.1.1-py3-none-any.whl
Size	7.0 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`301f739af54206a9d6dc87c17b43a682cced8840f607b1c95254971e67e79181`
BLAKE2b-256 checksum How to use checksums	`4f4f7672003c12b4a6145aa7b4d0e6559ad4b27c02a6301b6254311a5a5026a4`
Upload date	Apr 28, 2020
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/3.1.1 pkginfo/1.5.0.1 requests/2.11.1 setuptools/45.1.0 requests-toolbelt/0.9.1 tqdm/4.31.0 CPython/3.7.7`