Pruning and other network surgery for trained Keras models.
Switch branches/tags
Nothing to show
Clone or download
BenWhetton Fix `get_apoz` failing to identify image generator
Also fix a bug in calculating the apoz when layers are re-used.
Latest commit 459e8df Oct 23, 2018

README.md

Keras-surgeon

Introduction

Keras-surgeon provides simple methods for modifying trained Keras models. The following functionality is currently implemented:

  • delete neurons/channels from layers
  • delete layers
  • insert layers
  • replace layers

Keras-surgeon is compatible with any model architecture. Any number of layers can be modified in a single traversal of the network.

These kinds of modifications are sometimes known as network surgery which inspired the name of this package.

The operations module contains simple methods to perform network surgery on a single layer within a model.
Example usage:

from kerassurgeon.operations import delete_layer, insert_layer, delete_channels
# delete layer_1 from a model
model = delete_layer(model, layer_1)
# insert new_layer_1 before layer_2 in a model
model = insert_layer(model, layer_2, new_layer_3)
# delete channels 0, 4 and 67 from layer_2 in model
model = delete_channels(model, layer_2, [0,4,67])

The Surgeon class enables many modifications to be performed in a single operation.
Example usage:

# delete channels 2, 6 and 8 from layer_1 and insert new_layer_1 before 
# layer_2 in a model
from kerassurgeon import Surgeon
surgeon = Surgeon(model)
surgeon.add_job('delete_channels', model, layer_1, channels=[2, 6, 8])
surgeon.add_job('insert_layer', model, layer_2, new_layer=new_layer_1)
new_model = surgeon.operate()

The identify module contains methods to identify which channels to prune.

Documentation

The docstrings and this file contain all of the documentation. Standalone documentation may be added in the future.

Motivation

This project was motivated by my interest in deep learning and desire to experiment with some of the pruning methods I have read about in the research literature. I could not find an easy way to prune neurons from Keras models.

I hope I have created something which will be useful to others.

Installation

pip install kerassurgeon

Examples:

Examples are in kerassurgeon.examples.
Both examples identify which neurons to prune using the method described in Hu et al. (2016): those which have the highest Average Percentage of Zeros (APoZ).
Neither example is particularly good at demonstrating the benefits of pruning but they show how Keras-surgeon can be used.
I would welcome any good examples from other users.

Pruning Lenet trained on MNIST:

lenet_minst is a very simple example showing the effects of deleting channels from a simple Lenet style network trained on MNIST. It demonstrates using the simple methods from kerasurgeon.operations.

Inception V3 fine-tuned on flowers data-set:

This example shows how to delete channels from many layers simultaneously using the Surgeon Class.
It is in two parts:
inception_flowers_tune shows how to fine-tune the Inception V3 model on a small flowers data set (based on a combination of Tensorflow tutorial and Keras blog post).
inception_flowers_prune demonstrates deleting channels from many layers simultaneously using the Surgeon Class.

Limitations:

Only python 3 is currently supported. Only python 3.5 has been tested.
The following layers are not fully supported; delete_channels might not work on models containing these layers (it depends if they are affected by the operation):

  • Lambda
  • SeparableConv2D
  • Conv2DTranspose
  • LocallyConnected1D
  • LocallyConnected2D
  • TimeDistributed
  • Bidirectional
  • Dot
  • PReLU

Recurrent layers’ sequence length must be defined.
The model’s input shape must be defined.

Future improvements:

Architecture:

Investigate more efficient ways of modifying a layer in the middle of a model without re-building the whole network.

Performance:

This package has not yet been optimised for performance. It can certainly be improved.

Tests:

Write unit tests for the utility functions.
This package pretty tightly coupled with Keras which makes unit testing difficult. Some component tests have been written but it needs more work.

Examples

Write better examples.