ashnkumar/sketch-code

SketchCode

Generating HTML Code from a hand-drawn wireframe

Built in 2018, preserved as-is. SketchCode came out of the Insight AI Fellowship, back when solving this meant training a model specifically for it. The technique of the moment was image captioning — a CNN encoder feeding an LSTM decoder — and the bet was that the same architecture would work if you treated HTML as the caption. It did, on wireframes close to the 1,700-image synthetic dataset it was trained on. The code targets TensorFlow 1.x and Keras of that era and is no longer maintained; it stands as a record of what this problem took before general-purpose vision-language models started doing it zero-shot.

SketchCode is a deep learning model that takes hand-drawn web mockups and converts them into working HTML code. It uses an image captioning architecture to generate its HTML markup from hand-drawn website wireframes.

For more information, check out this post: Automating front-end development with deep learning

This project builds on the synthetically generated dataset and model architecture from pix2code by Tony Beltramelli and the Design Mockups project from Emil Wallner.

Note: This project is meant as a proof-of-concept; the model isn't (yet) built to generalize to the variability of sketches seen in actual wireframes, and thus its performance relies on wireframes resembling the core dataset.

Setup

Prerequisites

• Python 3 (not compatible with python 2)
• pip

Install dependencies

pip install -r requirements.txt

Example Usage

Download the data and pretrained weights:

# Getting the data, 1,700 images, 342mb
git clone https://github.com/ashnkumar/sketch-code.git
cd sketch-code
cd scripts

# Get the data and pretrained weights
sh get_data.sh
sh get_pretrained_model.sh

Converting an example drawn image into HTML code, using pretrained weights:

cd src

python convert_single_image.py --png_path ../examples/drawn_example1.png \
      --output_folder ./generated_html \
      --model_json_file ../bin/model_json.json \
      --model_weights_file ../bin/weights.h5

General Usage

Converting a single image into HTML code, using weights:

cd src

python convert_single_image.py --png_path {path/to/img.png} \
      --output_folder {folder/to/output/html} \
      --model_json_file {path/to/model/json_file.json} \
      --model_weights_file {path/to/model/weights.h5}

Converting a batch of images in a folder to HTML:

cd src

python convert_batch_of_images.py --pngs_path {path/to/folder/with/pngs} \
      --output_folder {folder/to/output/html} \
      --model_json_file {path/to/model/json_file.json} \
      --model_weights_file {path/to/model/weights.h5}

Train the model:

cd src

# training from scratch
# <augment_training_data> adds Keras ImageDataGenerator augmentation for training images
python train.py --data_input_path {path/to/folder/with/pngs/guis} \
      --validation_split 0.2 \
      --epochs 10 \
      --model_output_path {path/to/output/model}
      --augment_training_data 1

# training starting with pretrained model
python train.py --data_input_path {path/to/folder/with/pngs/guis} \
      --validation_split 0.2 \
      --epochs 10 \
      --model_output_path {path/to/output/model} \
      --model_json_file ../bin/model_json.json \
      --model_weights_file ../bin/pretrained_weights.h5 \
      --augment_training_data 1

Evalute the generated prediction using the BLEU score

cd src

# evaluate single GUI prediction
python evaluate_single_gui.py --original_gui_filepath  {path/to/original/gui/file} \
      --predicted_gui_filepath {path/to/predicted/gui/file}

# training starting with pretrained model
python evaluate_batch_guis.py --original_guis_filepath  {path/to/folder/with/original/guis} \
      --predicted_guis_filepath {path/to/folder/with/predicted/guis}

License

MIT — see LICENSE.

Portions of this project derive from pix2code, which is licensed under the Apache License, Version 2.0. See NOTICE for the required attributions and the terms that apply to those portions.