Skip to content

Latest commit

 

History

History
133 lines (68 loc) · 4.28 KB

README.md

File metadata and controls

133 lines (68 loc) · 4.28 KB
Raw

CycleGAN-Photo-to-Monet

PROJECT OVERVIEW

Convert photo images into Monet-like paintings using CycleGAN network

Project goals:

  1. Build, train and tune CycleGAN neural network for style transfer
  2. Create web app using Gradio
  3. Create demo
  4. Create Docker image

Contents

Training

  • Training dataset contents 300 Monet paintings and 7000 photos size of 256x256 px

  • During preprocessing input images were normalized

  • As a model CycleGAN neural network was used for training

  • CycleGAN network was constructed from several blocks:

    1. Generator Photo that generates photo-like images from input

    2. Generator Monet that generates Monet-like images from input

    3. Discriminator photo - it should predict if input photo image was generated - then output '0' or it is real - in that case output '1'

    4. Discriminator Monet - it should predict if input Monet image was generated - then output '0' or it is real - in that case output '1'

  • The main puprpose of generators is to produce monet style painting from photo input for monet-generator and vice verse for photo-generator

  • The main purpose of discriminator is to correctly identify if input image was generated or is it real

  • Scheme of CycleGAN training:

  • After each training epoch losses were calculated:

    1. Identity Loss for photo-generator and monet-generator

    2. Adversarial Loss for photo-generator and monet-generator

    3. Cycle Loss for photo-generator and monet-generator

    4. Dicriminator Loss for photo-discriminator and monet-discriminator

  • Neural Network was tuned with following hyper-parameters:

    1. Learning rate (final choise - 0.01 with decay to 0.001)

    2. Optimizer type (final choice - Adam)

    3. Optimizer momentums (final choice - first momentum=0.5, second = 0.999)

    4. Number of epoches (final choice - 100)

  • More information about data analysis and training process can be found in notebook: cylcegan-training.ipynb

  • Trained model: model.ckpt

Results

  • On the following plot you can see values of generators and discriminators losses during training:

  • Example of generator prediction on photo images after training:

Web App

  • To run Gradio web app type the following command:

    python3 app.py

  • Go to '''http://0.0.0.0:7000''' address in browser

  • Upload image or choose from examples and press 'Submit' to get the prediction

Docker

  • Install Docker: docker

  • Download Docker image from repository: docker pull plasticglass/painter:latest

  • (Optionaly) Build Docker image: docker build -t painter -f Docker/Dockerfile

  • Run following command to execute docker container:

    docker run -it -p 7000:7000 plasticglass/painter then go to http://0.0.0.0:7000

Modules and Tools

Web Development:

Docker | Gradio

Python-CNN:

Python | Pandas | Numpy | Pillow | Torch | Torchvision