PreTrainingChain 0.0.2

Scalable, configurable and Pre-training DNN using chainer

Introduction

============



Extension of chainer. ChainList for the purpose of network

scalability/congirablity/Pre-training executablity for deep leaning.

(You need to get deep learning framework "chainer" from

http://chainer.org/)

Detail of this project is here.

https://github.com/fukatani/PreTrainingChain



feature:

========



1) You can define network structure by list or tuple such as [784, 250, 200, 160, 10].
--------

This feature accelerate your deep network development. If you call this

class by AbstractChain([784, 250, 200, 160, 10]), you can generate

ChainList-> (F.Linear(784, 250), F.Linear(250, 200), F.Linear(200, 160),

F.Linear(160, 10)) You can change network structure without any hard

coding.



2) Pre-training executable.

--------


You can execute pre-training only by calling

AbstractChain.pre\_training(train\_data). Pretraining is executed by

using Bengio method. (http://arxiv.org/pdf/1206.5538.pdf) If length of

train\_Data is zero, Pre-training is skipped.