simo23 / conv2d_transpose

How Tensorflow's conv2d_transpose works

How Tensorflow's conv2d_transpose works

This is a simple tutorial to show how Tensorflow's conv2d_transpose works. I obtain the conv2d_transpose() output using conv2d_backprop_input() and conv2d() functions by performing deconvolution from the activations of the first convolutional layer. The results are that conv2d_transpose() transposes the filter's weights and flips them by 180 degrees.

I test this using a CNN on the MNIST dataset which obtains 99,2% test accuracy. The checkpoint of the already trained network is given.

The details of the CNN are:

1. Input layer: the inputs are MNIST images of 28x28x1
2. Conv1 layer: convolutional layer with 32 filters of 7x7, input = 1x28x28x1 , output = 1x28x28x32 , weights = 7x7x1x32
3. Conv2 layer: conv layer with 64 filters, input= 32x28x28x1 , output = 64x28x28x1 , weights = 3x3x32x64
4. FC layer with 1024 units
5. FC layer with 10 units
6. Output = 10 scalars used to classify into the 10 MNIST's classes

Requirements

I used Tensorflow 1.0 on Ubuntu 16.04, numpy 1.13.0, python 3.4

To use the code:

• Check requirements ( not sure if it will work with other versions )
• Download all the files and place them into a single folder
• Download the checkpoint from my drive: https://drive.google.com/file/d/0B6UbcmAnS9tCUUJaZDBJQ25kSWM/view?usp=sharing
• Extract the ckpt files and place them directly into the folder where the scrips are ( there should be 4 ckpt files in total )
• Do not copy the ckpt folder, place all the scripts and ckpt files into one folder
• Unzip the MNIST_data
• Run the test.py script
• An image should compare in the folder