Seung-Hun-Lee / DRANet

DRANet (CVPR 2021)

Pytorch implementation of paper:

Seunghun Lee, Sunghyun Cho, Sunghoon Im, "DRANet: Disentangling Representation and Adaptation Networks for Unsupervised Cross-Domain Adaptation", CVPR (2021)

Requirements

Pytorch 1.8.0
CUDA 11.1
python 3.8.10
numpy 1.21.0
scipy 1.7.1
tensorboardX
prettytable

Data Preparation

Download MNIST-M, Cityscapes, GTA5

Folder Structure of Datasets

├── data
      ├── MNIST
      ├── USPS
      ├── mnist_m
            ├── mnist_m_train
                      ├── *.png
            ├── mnist_m_test
                      ├── *.png
            ├── mnist_m_train_labels.txt
            ├── mnist_m_test_labels.txt
      ├── Cityscapes
            ├── GT
                   ├── train
                   ├── val
                   ├── test
            ├── Images
                   ├── train
                   ├── val
                   ├── test
      ├── GTA5
            ├── GT
                   ├── 01_labels
                   ├── 02_labels
                   ├── ...
            ├── Images
                   ├── 01_images
                   ├── 02_images
                   ├── ...
      
├── data_list
      ├── Cityscapes
              ├── train_imgs.txt
              ├── val_imgs.txt
              ├── train_labels.txt
              ├── val_labels.txt
      ├── GTA5
              ├── train_imgs.txt
              ├── train_labels.txt

Train

You must input the task(clf or seg), datasets(M, MM, U, G, C), and experiment name.

python train.py -T [task] -D [datasets] --ex [experiment_name]
example) python train.py -T clf -D M MM --ex M2MM

Test

Input the same experiment_name that you trained and specific iteration.

python test.py -T [task] -D [datasets] --ex [experiment_name (that you trained)] --load_step [specific iteration]
example) python test.py -T clf -D M MM --ex M2MM --load_step 100000

Tensorboard

You can see all the results of each experiment on tensorboard.

CUDA_VISIBLE_DEVICES=-1 tensorboard --logdir tensorboard --bind_all