Unet is the most well-known segmentation network for segmentation. It was first designed to adapt the biomedical image domain. In this project, we will test the performance of UNet with an urban city datasets - Cityscapes
All documents related to this repo can be found here:
pip install -r requirements.txt
-
Prepare your dataset
- First, move into .notebook directory and find data_prepare.ipynb notebook, run first 4 cells to download the Cityscapes dataset. You'll need to have an account. Sign up for a new account at Cityscapes's Homepage
- After this, it will automatically download and split the datasets into train, val and test subsets. The dataset's size is about 11GB.
-
This project will be updated for CLI interaction, but now, if you want to change the model architechture, hyperparameters and so on ... Please modify it in python code train.py
-
Run
-
Start training from scratch:
python train.py
-
This project support several model architechture:
- FCN , UNet , UNet with ResNet encoder , PSPNet, DeepLabv3.
-
usage: test.py
The logs during the training will be stored, and you can visualize it using TensorBoard by running this command:
tensorboard --logdir ~/saved/<project_name>
# specify a port 8080
tensorboard --logdir ~/saved/<project_name> --port 8080