This repo is an implementation of BiSeNet in Keras on the Cityscapes dataset. But it still has some differences compared with the author's.The code is currently available and will fix issues in the code later.
The context_model mainly contains ResNet18,ResNet50 and Xception which you can use the pretrained model provided by Keras-Model.
You can initialize the model weights using the Keras pretrained model weights.
Here I put my trained BiSeNet-Xception weights on the Google Drive which can achieve 64.3% mIoU on the Cityscapes validation set.