CityLayout

👆 CityLayout is a generative city 2d layout workflow, you can custom your OSM (openstreetmap) dataset by using provided scripts automatically. The dumped raw data will be stored as GeoJSON format and after preprocessing, it will be saved as one-hot encoded numpy array (eg. 0: road, 1: building, 2: water, 3: park, 4: etc). You can training with your custom dataset or just use pretrained model to generate city layout.

Todo

• Release code
• Release dataset preprocess script
• Release pretrain ckpt
• Release dataset
• Webui demo

Setup environment

# conda env
conda create -f environment.yml
# pip install
pip install -r requirements.txt

Download pretrain ckpt

You can download our pretrained model from here You can find three different ckpt folders, the details are as follows:

ckpt	Description
CityGen/best_ckpt.pth or latest_ckpt.pth	The pretrained model under CityGen's setting
normal/best_ckpt.pth or latest_ckpt.pth	The pretrained model for citylayout random generation
Completion/best_ckpt.pth or latest_ckpt.pth	The pretrained model for citylayout completion

All the ckpts mentioned above are trained with the same dataset setting , the training/sampling citylayout include building, road, and waterbody.

After downloading the ckpt, create soft link to ckpts folder.

ln -s /path/to/ckpt ckpts

Download dataset (optional & not released yet)

We provide our used dataset for training and evaluation, you can download it from #### and create soft link to data folder.

ln -s /path/to/data data

The data dir may look like this:

data
├── train
│   ├── cityname-0-0.h5
│   ├── cityname-0-1.h5
│   ├── cityname-0-2.h5
...
├── val
│   ├── cityname-0-0.h5
...
├── test
│   ├── cityname-0-0.h5
...

The provided data has been preprocessed and saved as h5 format, you can directly use it for training and evaluation.

Download dataset and preprocess

This tutorial will help you create yout own osm training data from scratch

First: get target position (lat, lon) by runing script get_locations

the query list is stored in config/data/city_landmark.yaml

# get target position 
python scripts/osm/get_locations.py

Second: download osm data by runing script get_city_osm.py

The default config for downloading osm data is stored in config/data/osm_cities.yaml, including some path like input (city coordinates), output (osm data), and some other parameters like citylayout patch size(pixel), etc.

# download osm data
python scripts/osm/get_city_osm.py

Third: preprocess osm data by runing script preprocess.py

The default config for preprocessing osm data is stored in config/data/preprocess.yaml, which provides detailed osm customizing parameters like selected tags, geo coordinate, image size, etc. Preprocessed data will be stored as one-hot encoded numpy array with h5 format under boolean type.

# preprocess osm data
python scripts/osm/preprocess.py

Train model

TBD

Generate city layout

You can generate citylayout from gaussian noise by using normal ckpt, or you can generate citylayout from existing citylayout by using citygen ckpt with inpainting or outpainting mode, lastly, you can refine the citylayout by using completion ckpt.

Generate city layout from gaussian noise

# random sample
python src/train_lightning.py -cn=refactoring_sample
# with refiner 
python src/train_lightning.py -cn=refactoring_sample CityDM.Test.use_completion=True

Generate city layout from existing city layout

# inpainting
python src/train_lightning.py -cn=refactoring_sample_citygen CityDM.Test.sample_type=Inpainting
# outpainting
python src/train_lightning.py -cn=refactoring_sample_citygen CityDM.Test.sample_type=Outpainting

for compuational cluster, you can use the following command to generate city layout

# random sample
sbatch scripts/run/sample_normal.sh