NN_Project

This is the course project for CSCE636 Neural Network.

Directories: Models: this folder contains the trained models for Van Gogh, Renoir and Monet. It also contains the graph of these three models.

src: this folder contains the code that is needed while applying models.

Styles: this folder contains three original masterpieces and their doodles. The trained models are trained by them. The user can also use them to train new models.

Train: this folder contains all files that are needed to train a new model.

GUI

Simply run

python GUI.py

Here is the demo of GUI.

Note: the GPU server doesn't support GUI, so I haven't tried GUI.

Stylize Doodle

Use trained model to proecess the doodle.

CUDA_VISIBLE_DEVICES=2,3 python apply.py --colors model_color.npy --target_mask target_maskfile --model model_name.t7

Example:

CUDA_VISIBLE_DEVICES=2,3 python apply.py --colors Models/VanGogh.hdf5_colors.npy --target_mask target_mask.png --model Models/VanGogh.t7

Training

Generate datasets

cd Train
python generate.py --n_jobs 30 --n_colors 4 --style_image style_image_path --style_mask style_image_mask_path --out_hdf5 dataset_path

Example:

python generate.py --n_jobs 30 --n_colors 4 --style_image Styles/Monet.png --style_mask Styles/Monet_mask.png --out_hdf5 Monet.hdf5

Train the model

You need to download VGG-19 recognition network.

cd data/pretrained && bash download_models.sh && cd ../..

Then train the model

CUDA_VISIBLE_DEVICES=2,3 th feedforward_neural_doodle.lua -model_name skip_noise_4 -masks_hdf5 dataset_path -batch_size 4 -num_mask_noise_times 0 -num_noise_channels 0 -learning_rate 1e-1 -half false

Example:

CUDA_VISIBLE_DEVICES=2,3 th feedforward_neural_doodle.lua -model_name skip_noise_4 -masks_hdf5 Monet.hdf5 -batch_size 4 -num_mask_noise_times 0 -num_noise_channels 0 -learning_rate 1e-1 -half false

Here is the demo of Training.

Prerequisites

• torch
  • torch7
• python
  • sklearn
  • skimage
  • numpy
  • scipy
  • h5py
  • joblib
  • tkinter

A good guide on installation can be found here.

The code is tested by Python2.7 and the lasted conda.

Credits

The code is based on Dmitry Ulyanov's code.