Pytorch Implementation of the paper "Cross-domain Correspondence Learning for Exemplar-based Image Translation" (CVPR 2020 oral).
20200525: Training code for deepfashion complete. Due to the memory limitations, I employed the following conversions:
- Disable the non-local layer, as the memory cost is infeasible on common hardware. If the original paper is telling the truth that the non-lacal layer works on (128-128-256) tensors, then each attention matrix would contain 128^4 elements (which takes 1GB).
- Shrink the correspondence map size from 64 to 32, leading to 4x memory save on dense correspondence matrices.
- Shrink the base number of filters from 64 to 16.
The truncated model barely fits in a 12GB GTX Titan X card, but the performance would not be the same.
- Ubuntu/CentOS
- Pytorch 1.0+
- opencv-python
- tqdm
- Prepare dataset
- Implement the network
- Implement the loss functions
- Implement the trainer
- Training on DeepFashion
- Adjust network architecture to satisfy a single 16 GB GPU.
- Training for other tasks
Just follow the routine in the PATN repo
The pretrained model for human pose transfer task: TO BE RELEASED
run python train.py.
If you find this repo useful for your research, don't forget to cite the original paper:
@article{Zhang2020CrossdomainCL,
title={Cross-domain Correspondence Learning for Exemplar-based Image Translation},
author={Pan Zhang and Bo Zhang and Dong Chen and Lu Yuan and Fang Wen},
journal={ArXiv},
year={2020},
volume={abs/2004.05571}
}
TODO.