mkmenta / domain_adapt_segm

Implementation of my Master Thesis "Learning to adapt class-specific features across domains for semantic segmentation".

Learning to adapt class-specific features across domains for semantic segmentation

Master thesis available here.

Requirements

Implemented using PyTorch v0.4.0 and Python 3.

Datasets:

• Get MNIST from yann.lecun.com
• Get MNIST-M from yaroslav.ganin.net
• Get ThinMNIST from my GitHub

Training commands

Apart from the following arguments, you will need to specify an exp_name, mnist_path, mnist_m_path and mnist_thin_path.

Specify a GPU to run the code adding CUDA_VISIBLE_DEVICES=X before the command.

Example of training command for the FCN Segmenter baseline:

python train_segm_baseline.py

Example of training command for the SGAN-S baseline:

python main.py --mode train

Example of training command for the SGAN-S + Uncond. baseline:

python main.py --mode train --da_type uncond --df_num_down 2 --lambda_fdom 1 --lambda_frf 1

Example of training command for the SGAN-S + In. Cond.:

python main.py --mode train --da_type input_cond --df_num_up 2 --df_num_down 4 --lambda_frf 1

Example of training command for the SGAN-S + Out. Cond.:

python main.py --mode train --da_type output_cond --df_num_up 2 --lambda_frf 1 --lambda_fdom 1