Instance-wise Hard Negative Example Generation for Contrastive Learning in Unpaired Image-to-Image Translation (NEGCUT)

We provide our PyTorch implementation of Instance-wise Hard Negative Example Generation for Contrastive Learning in Unpaired Image-to-Image Translation (NEGCUT). In the paper, we identify that the negative examples play a critical role in the performance of contrastive learning for image-to-image translation. We train a generator to generate negative examples online through adversarial learning to enhance the performance of contrastive learning in unpaired image-to-image translation. Compared to CUT, our model achieves superior performances on three benchmark datasets with the same inference speed.

Example Results

• Cat->Dog:

• Horse->Zebra:

• Cityscapes:

Prerequisites

• Linux
• Python 3
• CPU or NVIDIA GPU + CUDA CuDNN

Getting started

• Install PyTorch and other dependencies (e.g., torchvision, visdom, dominate, gputil).

  For pip users, please type the command pip install -r requirements.txt.

  For Conda users, you can create a new Conda environment using conda env create -f environment.yml.

NEGCUT Training and Test

• Download the cityscapes dataset.

bash ./datasets/download_cut_dataset.sh cityscapes

Preprocess the cityscapes dataset with the scripts datasets/prepare_cityscapes_dataset.py.

python prepare_cityscapes_dataset.py --gitFine_dir ./gtFine/ --leftImg8bit_dir ./leftImg8bit --output_dir ./datasets/cityscapes/

The dataset will be saved at ./datasets/cityscapes/.

• To view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097.

• Train the NEGCUT model:

python train.py --dataroot ./datasets/horse2zebra --name CITY_NEGCUT --NEGCUT_mode NEGCUT --model negcut

The checkpoints will be stored at ./checkpoints/CITY_NEGCUT/web.

• Test the CUT model:

python test.py --dataroot ./datasets/cityscapes --name CITY_NEGCUT --NEGCUT_mode NEGCUT --model negcut --phase test

The test results will be saved to a html file here: ./results/cityscapes/latest_train/index.html.

Apply a pre-trained NEGCUT model and evaluate

The pretrained models can be downloaded at [checkpoints].

• To evaluate the model, first generate the test results:

python test.py --dataroot ./datasets --name MODEL_NAME --NEGCUT_mode NEGCUT --model negcut --phase test

• To calucate FID metric, you should clone pytorch-fid and run:

python fid_score.py GENERATED_IMAGES_DIR REAL_IMAGE_DIR

• To calcuate mAP, PixAcc and ClsAcc metrics, you should clone drn and apply the pre-trained drn-d-22 model at assets/drn/drn_d_22.pth.tar for the resolution of 128x256 [checkpoints]:

python segment.py test -d <data_folder> -c 19 --arch drn_d_22 --batch-size 1 --resume assets/drn/drn_d_22.pth.tar --phase val --with-gt

To evaluate your generated images, you need to arrange your generated results like the cityscapes dataset, refer drn/datasets/cityscapes/prepare_data.py for more details.

Datasets

Refer Datasets to learn more details about datasets used and how to create your own datasets

Acknowledge

Our code is developed based on CUT. We also thank pytorch-fid for FID computation, drn for mIoU computation.