Coarse-to-Fine Generative Modeling for Graphic Layouts
Authors: Zhaoyun Jiang, Shizhao Sun, Jihua Zhu, Jian-Guang Lou, Dongmei Zhang1096-1103
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach qualitatively and quantitatively on UI layouts from RICO (Deka et al. 2017) and document layouts from Pub Lay Net (Zhong, Tang, and Yepes 2019). Experiments show that our approach outperforms existing approaches, especially on complex layouts which have many elements and complicated element arrangements. |
| Researcher Affiliation | Collaboration | 1 School of Computer Science and Technology, Xi an Jiaotong University, Xi an, China 2 Microsoft Research Asia, Beijing, China 3 School of Software Engineering, Xi an Jiaotong University, Xi an, China |
| Pseudocode | No | The paper describes the model architecture and training process in text and with diagrams (Figure 2), but does not provide any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about making the source code available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We evaluate our method on the following publicly available datasets, which are widely used in recent studies about graphic layout generation. RICO (Deka et al. 2017) contains 66K+ mobile app UI layouts... Pub Lay Net (Zhong, Tang, and Yepes 2019) contains 360K+ document layouts... |
| Dataset Splits | Yes | In total, we get 54K screenshot data for training and 6K data for validation. Pub Lay Net...finally get 300K data for training and 33K for validation. |
| Hardware Specification | Yes | All models are trained for 300 epochs on RICO and 100 epochs on Pub Lay Net, with batch size of 256 on two V100 GPUs. |
| Software Dependencies | No | The paper states, 'Our approach is implemented by Py Torch,' but it does not provide specific version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | For Transformer blocks, we stack 4 layers with a representation size of 512 and a feed-forward representation size of 1024, and use multi-head attentions with 4 heads. We use Adam optimizer (Kingma and Ba 2014) with initial learning rate 10 3 reduced by a factor of 0.8. The dropout rate of transformer blocks is set to 0.1. All models are trained for 300 epochs on RICO and 100 epochs on Pub Lay Net, with batch size of 256 on two V100 GPUs. |