Code and models for urban fabric generation using Pix2pix and GauGAN
该项目研究深度学习相关技术在城市肌理生成上的应用可能性，使用Pix2pix和GauGAN两种图像转换模型对肌理成对数据进行学习，训练好的模型可以基于土地利用信息（不同的颜色标注）来生成相应的肌理。在项目中，我尝试了一种基于自编码器的特征提取方法，来比较不同模型生成效果上的差异，以此调制出更适合于肌理生成的模型超参数和模型架构，最后将模型应用于步进式条件生成和城市尺度肌理生成。
This project researches on applications of deep learning technologies in urban texture generation, using Pix2pix and GauGAN models to learn urban fabric paired data. The trained model can generate the corresponding urban fabric based on land use condition(different color labeling). In the project, I tried a feature extraction method based on Autoencoder to compare the differences in the generation effect of different models, so as to modulate the model hyperparameters and model architecture which are more suitable for urban fabric generation. Finally, pretained models can ben applied to step-by-step conditional generation and city-scale fabric generation.
该项目使用Pytorch 1.2实现（1.0版本推测应该都可以），运行项目代码需要在你的Python环境中安装相应的库，可通过以下命令实现：
The project is implemented using Pytorch 1.2 (version 1.0 should work). To run the code, you need to install the required libraries in your Python environment first, which can be done by typing this one line command:
pip install -r requirements.txt
文件中已经放入城市肌理数据集和雄安西区容西片区土地利用规划简化图，默认参数下,4G显存的电脑即可完整运行代码来进行模型训练和生成（前提是已安装CUDA和相应的cudnn），如果配置允许，可适当调大Batch size或者调整网络架构（架构调整可通过options参数界面来实现），以达到更好的模型训练效果。项目所用模型在16G Tesla V100显卡上训练200 epochs得到，Batch size为12。 由于预训练模型和完整数据集较大，因此文件中不包含项目最终训练完成的模型，城市肌理数据集也仅提供一部分用于测试（500 training data 和 100 testing data）。如有这两方面急切需求，可通过微信联系我：18482185419
Dataset of urban fabric and the simplified image of land use planning in Rongxi Xiongan have been put into this project. Under the default parameters, code can be run on a computer with 4G GPU for model training and generation(with CUDA and the corresponding cudnn installed). if the configuration allows, you can appropriately adjust the Batch size or adjust the network architecture (architecture adjustment can be achieved through options module) to achieve better model training results. The model used in the project was trained on a 16G Tesla V100 GPU for 200 epochs, Batch size is 12. Due to pre-training model and complete dataset are too large, the final trained model is not included in the project and only part of urban fabric dataset is provided for code testing(500 training data and 100 testing data). If you have an urgent need in these two aspects, you can contact me through my Wechat: 18482185419