emilwallner / Screenshot To Code

A detailed tutorial covering the code in this repository: Turning design mockups into code with deep learning.

Plug: I write about learning Machine learning online and independent research 👉 Enjoy!

The neural network is built in three iterations. Starting with a Hello World version, followed by the main neural network layers, and ending by training it to generalize.

The models are based on Tony Beltramelli's pix2code, and inspired by Airbnb's sketching interfaces, and Harvard's im2markup.

Note: only the Bootstrap version can generalize on new design mock-ups. It uses 16 domain-specific tokens which are translated into HTML/CSS. It has a 97% accuracy. The best model uses a GRU instead of an LSTM. This version can be trained on a few GPUs. The raw HTML version has potential to generalize, but is still unproven and requires a significant amount of GPUs to train. The current model is also trained on a homogeneous and small dataset, thus it's hard to tell how well it behaves on more complex layouts.

A quick overview of the process:

1) Give a design image to the trained neural network

2) The neural network converts the image into HTML markup

3) Rendered output

Installation

FloydHub

Click this button to open a Workspace on FloydHub where you will find the same environment and dataset used for the Bootstrap version. You can also find the trained models for testing.

Local

pip install keras tensorflow pillow h5py jupyter

git clone https://github.com/emilwallner/Screenshot-to-code.git
cd Screenshot-to-code/
jupyter notebook

Go do the desired notebook, files that end with '.ipynb'. To run the model, go to the menu then click on Cell > Run all

The final version, the Bootstrap version, is prepared with a small set to test run the model. If you want to try it with all the data, you need to download the data here: https://www.floydhub.com/emilwallner/datasets/imagetocode, and specify the correct dir_name.

Folder structure

  |  |-Bootstrap                           #The Bootstrap version
  |  |  |-compiler                         #A compiler to turn the tokens to HTML/CSS (by pix2code)
  |  |  |-resources											
  |  |  |  |-eval_light                    #10 test images and markup
  |  |-Hello_world                         #The Hello World version
  |  |-HTML                                #The HTML version
  |  |  |-Resources_for_index_file         #CSS,images and scripts to test index.html file
  |  |  |-html                             #HTML files to train it on
  |  |  |-images                           #Screenshots for training
  |-readme_images                          #Images for the readme page

Hello World

HTML

Bootstrap

Model weights

• Bootstrap (The pre-trained model uses GRUs instead of LSTMs)
• HTML

Acknowledgments

• Thanks to IBM for donating computing power through their PowerAI platform
• The code is largely influenced by Tony Beltramelli's pix2code paper. Code Paper
• The structure and some of the functions are from Jason Brownlee's excellent tutorial