tf_fast_neural_style

A fast neural style transfer network coded in Tensorflow

This project is based on the Stanford University Paper: https://arxiv.org/abs/1603.08155

Requirements:

Python 3.5
Tensorflow 1.3
Scipy
Pretrained VGG19 Network: http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat

Training:

Requires a TFRecords file of jpeg images
Training on PASCAL VOC 2007 Dataset is recommended, will resize to 256 x 256 pixels

Running the program:

The program uses the Python Argument Parser Module and includes help documentation that lists the arguments such as filepaths and batchsize

Checkpoints and Visualization:

Utilizes Tensorboard to track training loss every 100 Iterations and to visualize the computation graph
Checkpoints are saved every 100 Iterations for easy testing of current results

Styling Images:

After a checkpoint is generated by train.py, styleimage.py can be run to transform the desired image into its styled counterpart

This was a great project to learn Tensorflow and build my own network from scratch! Enjoy!