ASM

Introduction

This is an official implementation for our NeurIPS 2020 paper: Adversarial Style Mining for One-Shot Unsupervised Domain Adaptation. In this paper, we aim at the problem named One-Shot Unsupervised Domain Adaptation. Unlike traditional Unsupervised Domain Adaptation, it assumes that only one unlabeled target sample can be available when learning to adapt.

Usage

Prerequisites

• Python 3.6
• GPU Memory >= 32G

Download ImageNet-pretained DeepLab:

• Download DeepLab_resnet_pretrained_init-f81d91e8.pth and put it under pretrained/.

Download Pretained RAIN

• Download vgg_normalized.pth/decoder_iter_160000.pth/fc_encoder_iter_160000.pth/fc_decoder_iter_160000.pth and put them under pretrained/.

Download DataSets

• Download GTA5
• Download Cityscapes

Modify data path to your own

• https://github.com/RoyalVane/ASM/blob/1dac4bfc702da7aaff342683cad628b73807ab2e/ASM/ASM_train.py#L67
• https://github.com/RoyalVane/ASM/blob/1dac4bfc702da7aaff342683cad628b73807ab2e/ASM/ASM_train.py#L83
• https://github.com/RoyalVane/ASM/blob/1dac4bfc702da7aaff342683cad628b73807ab2e/ASM/ASM_evaluate.py#L14

Train

CUDA_VISIBLE_DEVICES=<gpu_id> python ASM_train.py --snapshot-dir ./snapshots/GTA2Cityscapes

Test

CUDA_VISIBLE_DEVICES==<gpu_id> python ASM_evaluate.py

Compute IOU

python ASM_IOU.py

Our Pretrained Model

We also provide our Pretrained ASM models for direct evaluation. These models are trained using 32G V100.

• The first model is consist with our reported IoU result in the paper. mIoU = 44.53:

• The second model is trained recently, whose performance is slightly higher than the paper. mIoU = 44.78:

Citation

• If you find this code useful, please consider citing

@inproceedings{Luo2020ASM,
title={Adversarial Style Mining for One-Shot Unsupervised Domain Adaptation},
  author={Luo, Yawei and Liu, Ping and Guan, Tao and Yu, Junqing and Yang, Yi},
  booktitle={Advances in Neural Information Processing Systems},
year={2020}
}

Related Works

• CLAN: One-shot UDA is a realistic but more challenging setting than UDA, which we tried to solve in our CVPR2019 oral paper "Taking A Closer Look at Domain Shift: Category-level Adversaries for Semantics Consistent Domain Adaptation".

• Copy and Paste GAN: RAIN is also employed as a strong data augmentation module in our CVPR2020 oral paper "Copy and Paste GAN: Face Hallucination from Shaded Thumbnails".