Adversarial Domain Adaptation for Real-time Semantic Segmentation

Final Project – "Advanced Machine Learning" 2021 - 2022 Course at PoliTO

Giulia D'Ascenzi, Patrizio de Girolamo, Carlos Rosero

PyTorch implementation of the algorithm implemented for the final project of the course "Advanced Machine Learning".

In the "report" folder can be consulted:

• The assignement of the work, filled with the results got.
• The final detailed report of the project.
• The slides of the presentation done by the group during the oral exam of February 2022.

Requirements: pytorch 0.4.1, python 3.6, torchinfo.

Datasets

The datasets used are subsets of the Cityscapes and GTA5 datasets. They can be downloaded here: data.zip

Overview

1. Fully supervised training with the target dataset.

The file FS_train.py can be used to train the network BiSeNet in a supervised way using the target dataset (Cityscapes).

Example:

python FS_train.py '--num_epochs', '50',
                    '--data', 'path/to/data/folder'
                            

2. Unsupervised domain adaptation

The file DA_train.py can be used to train the network BiSeNet using unsupervised adversarial learning for the GTA5 -> Cityscapes domain adaptation task.

The file model/discriminator.py contains the models of the two implemented discriminators: the Fully Convlutional discriminator and the Light Weight Discriminator.

The 'architectures' folder describes the architectures of the two discriminators and BiSeNet.

To use the Light Weight version set '--light_discriminator, True'.

In utils/FDA.py the Fourier Domain Adaptation transformation is implemented as described in the original code.

To use FDA set '--FDA, True' and select the desired value for beta with '--LB, 0.05'.

Example:

python DA_train.py  '--data-dir', 'path/to/data/folder',
                    '--num-steps', '50',
                    '--iter-size', '125'
                    '--light_discriminator', 'True',
                    '--FDA', 'True',
                    '--LB', '0.05'
                  

3. Test

The file eval.py contains the code to test a pretrained model.

Example:

python eval.py '--pretrained_model_path', 'path/to/pretrained/model',
               '--data', 'path/to/data/folder'