This repository contains the official code for
"LC-MSM: Language-Conditioned Masked Segmentation Model for Unsupervised Domain Adaptation"

• Ubuntu 20.04
• Python 3.8.5

• torch >= 1.8.0
• torchvision
• mmcv-full
• open-clip
• tqdm

To use this code, please first install the 'mmcv-full' by following the official guidelines (mmcv).

The requirements can be installed by the following command

pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html
pip install open_clip_torch

Citysacpes: Please, download leftImg8bit_trainvaltest.zip and trainvaltest.zip from here

GTA: Please download all images and label packages from here

SYNTHIA: Please download SYNTHIA-RAND-CITYSCAPES from here

please download the pre-trained weight for MiT-B5 via shell script

sh tools/download_checkpoints.sh

The following dataset is preprocessed in COCO format , but if you are using the raw json file you can preprocess with the script

python tools/convert_datasets/gta.py /your/path/
python tools/convert_datasets/cityscapes.py /your/path/
python tools/convert_datasets/synthia.py /your/path/

For convenience, provides and annotated config file of the adaptation model
Before training the data path in the datset config file should be modifed with your data path.

A training job can be launched using:

python run_experiment.py --config configs/daformer/gta2cs_uda_lc_msm.py

The checkpoint will be sevaed automatically in work_dirs, else you set a directory for it.

sh test.sh path/to/checkpoint/directory