Pix2HTML is my implementation of the Pix2Code paper. The goal of this project is to convert GUIs created by designers into valid HTML code using an end-to-end deep learning approach. This repository contains all the necessary code to train Pix2HTML using CNNs and LSTMs. You can find the training notebook on Kaggle and model parameters.
For text generation in the evaluation phase, I used:
- Greedy Search
- Beam Search (Beam = 3)
- Beam Search (Beam = 5)
and the BLEU scores were as follows:
Model | Greedy search | Beam Search, Beam = 3 | Beam Search, Beam = 5 |
---|---|---|---|
VGG + LSTMs | ~0.78 | ~0.80 | ~0.80 |
More experiments are coming.