UI2code is a tool to convert the UI design image into the skeleton code with deep learning methods. For example, given an input of Android UI design image, it can generate the corresponding Android XML skeleton code to developers. And developers just need to fill in some detailed attributes such as color, text. We believe that this tool can assist the front-end mobile developers to implement the UI design image from designers. Some brief introduction can be seen in http://tagreorder.appspot.com/ui2code.html
We have published this work in our ICSE'18 paper:
From UI Design Image to GUI Skeleton: A Neural Machine Translator to Bootstrap Mobile GUI Implementation
Chunyang Chen, Ting Su, Guozhu Meng, Zhenchang Xing, Yang Liu
The 40th International Conference on Software Engineering, Gothenburg, Sweden.
http://ccywch.github.io/chenchunyang.github.io/publication/ui2code.pdf
We adopt the UI testing to explore more than 5000 Android Apps crawled from the Google Play, and then take the screenshots as the UI design image and also collect the corresponding code. There are totally 29,887 screenshots (We have resize it as 300 * 200 and rotate them 90 degree for training), and corresponding source code. It can be downloaded in https://drive.google.com/open?id=17cRSdNPd7GoNuirE983S467kWbOLtiuw and decompressed it for using.
The project is written in Torch, and the evaluation needs Python.
The following lua libraries are required for the main model.
- tds
- class
- nn
- nngraph
- cunn
- cudnn
- cutorch
Note that currently it can only run in GPU
Perl is used for evaluating BLEU score.
We have prepared the dataset for training, validation and testing.
so we need to specify a data_base_dir storing the images, a label_path storing all labels (e.g., code sequence). Besides, we need to specify a data_path for the training (or test) data samples. The format of data_path shall look like:
<img_name1> <label_idx1>
<img_name2> <label_idx2>
<img_name3> <label_idx3>
...
where <label_idx> denotes the line index of the label (starting from 0).
We have stored our trained model in https://drive.google.com/open?id=10vStYFIwA2ofXSzaUWA4JSw69XH6U3b6, training data as train.lst, validation data as validate.lst, testing data as test_shuffle.lst.
The raw image data is in processedImage, and all code data in XMLsequence.txt with the vocabulary as xml_vocab.txt.
You can train the model with the following order:
th src/train.lua
-phase train -gpu_id 1
-model_dir model
-input_feed -prealloc
-data_base_dir data/processedImage/
-data_path data/train.lst
-val_data_path data/validate.lst
-label_path data/XMLsequence.lst
-vocab_file data/xml_vocab.txt
-max_num_tokens 100 -max_image_width 300 -max_image_height 200
-batch_size 20
-beam_size 5
-dropout 0.2
-num_epochs 10
In the default setting, the log file will be put to log.txt. The log file records the training and validation perplexities. model_dir speicifies where the models should be saved. Please fine-tune the parameter for your own purpose.
After training, you can load a model and use it to test on test dataset. We provide a model trained on the our data which can be downloaded together with the dataset in xxx
Now you can load the model and test on test set. Note that in order to output the predictions, a flag -visualize must be set.
You can use the following order to test it:
th src/train.lua -phase test -gpu_id 1
-load_model -model_dir model
-log_path log_test.txt
-visualize
-data_base_dir data/processedImage/
-data_path data/test_shuffle.lst
-label_path data/XMLsequence.lst
-output_dir results
-max_num_tokens 100 -max_image_width 300 -max_image_height 200
-batch_size 30 -beam_size 5
The test perplexity can be obtained after testing is finished. In order to evaluate the exact match and BLEU, the following command needs to be executed.
python evaluate/checkExeperiment.py
Note that if you change the directory of the results data, please change it accordingly in the file.
This work heavily depends on the https://github.com/harvardnlp/im2markup, thanks for their work.