VAE & CNN example of CIFAR10(Tensorflow 1.1x )

It is a tutorial of a VAE (Variational AutoEncoder) and CNN(Convolutional Neural Network) with Tensorflow and python.
For this tutorial, we use CIFAR-10.
That is a common benchmark in machine learning for image recognition.

CIFAR-10 is what ??

Like this.

---

---

The details are shown in the following URL.
http://www.cs.toronto.edu/~kriz/cifar.html

You can generate this figure by running cifar10_loader.py as follows;

$ python cifar10_loader.py

ということで，CIFAR-10でVAEによるreconstructやCNNによるclassificationをやるコードです．
Windows10 + Pycharm なら，tensorboardはwsl上で起動しましょう．
なぜなら，windowsがpermission周りにうるさかったりフォルダを手放さなかったりするから．

batch-normのコーディングや，cifar10のデータローダー，data argumentationの参考になるかな？ 画像を上図みたいに並べるのも，探すの結構難航したから，参考になるかも（難航したのは自分だけ．．．．？）．
就活でtensorflow書けるって言って歩ているのに，何も見せられるコードが無かったからコレ作りました．

Environment

• OS

  • Windows10 (+ wsl for tensorboard)
  • Ubuntu 16.04

• Python packages and version

  • Python 3.6.xxx
  • tensorflow (tensorflow-gpu) 1.1xxx
  • numpy 1.14.xxx
  • opencv-python 1.14.xxx
  • Pillow 5.xxx

Contents

This project contains the following codes.

Code	Explanation
main.py	Train the model of VAE (by train_vae() ) and CNN-classification (by train_classification() )
network.py	The model of VAE(3-layer-encoder +3-layer-decoder) & CNN(5layer+batch_norm)
cifar10_loader.py	Download CIFAR10-binary-data automatically and some techniques of data-argumentation are implemented.

For Quick Start

Run VAE model (Default)

$ python main.py

If you want to run CNN classification model, comment out train_vae() in main.py , and comment in train_classification().
By the tensorboard, you can see the loss (or generated images) transition.

$ cd <saved_dir>/VAEsample/<log_dir_name>
$ tensorboard --logdir=./ --port 6006