Pytorch implementation for EGFE: End-to-end Grouping of Fragmented Elements in GUI Designs with Multimodal Learning.

pip install -r requirements.txt

This is the Pytorch implementation of EGFE. It has been trained and tested on Linux (Ubuntu20 + Cuda 11.6 + Python 3.9 + Pytorch 1.13 + 4 * NVIDIA GeForce RTX 3090 GPU), and it can also work on Windows.

Clone this repo:

git clone [email protected]:test2975/EGFE.git
cd EGFE

cd sketch_dataset

• Generate design prototypes(*.sketch) to method format python scripts/convert_sketch.py --input [sketch_folder] --output ./data
• Generate train and test format index files python scripts/convert_config.py [config.json file path] ./data/output
• Since the scripts use sketchtool, it has to run on macOS.

• Contact the authours of this paper to request the training dataset, and modified the path of train_index_json and test_index_json in ./sketch_model/configs/config.py.

• Download the testing dataset, and have it in the 'dataset' folder.

• Start to train with

torchrun --nnodes 1 --nproc_per_node 4  main_ddp.py --batch_size 10 --lr 5e-4

• Download the pre-trained EGFE model, and put it in the './work_dir' folder.
• Start to test with

torchrun --nnodes 1 --nproc_per_node 4  main_ddp.py --evaluate --resume ./work_dir/set-wei-05-0849/checkpoints/latest.pth --batch_size 40

• Start to train with

torchrun --nnodes 1 --nproc_per_node 4  efficient_main.py --batch_size 4 --lr 5e-4

• Start to test with

torchrun --nnodes 1 --nproc_per_node 4  efficient_main.py --evaluate --resume ./work_dir/efficient_net/latest.pth --batch_size 8

• Start to train with

torchrun --nnodes 1 --nproc_per_node 4  vit_main.py --batch_size 4 --lr 5e-4

• Start to test with

torchrun --nnodes 1 --nproc_per_node 4  vit_main.py --evaluate --resume ./work_dir/vit/latest.pth --batch_size 8

• Start to train with

torchrun --nnodes 1 --nproc_per_node 4  sw_vit_main.py --batch_size 4 --lr 5e-4

• Start to test with

torchrun --nnodes 1 --nproc_per_node 4  sw_vit_main.py --evaluate --resume ./work_dir/swin/latest.pth --batch_size 8

Please refer to https://github.com/zjl12138/UILM/

Please refer to https://github.com/zjl12138/ULDGNN

We have released a UI component classifier (binary classification) for UIED trained on our dataset. It can be downloaded from pretrained model.

Please refer to the open-source code https://github.com/MulongXie/UIED

The implementations of EfficientNet, Vision Transformer, and Swin Transformer are based on the following GitHub Repositories. Thank for the works.

• EfficientNet: https://github.com/lukemelas/EfficientNet-PyTorch
• ViT: https://github.com/lucidrains/vit-pytorch
• Swin Transformer: https://github.com/microsoft/Swin-Transformer

You can download all the model checkpoints from pre-trained model.7z.