[CVPR 2020] GAN Compression: Efficient Architectures for Interactive Conditional GANs
Include the MUNIT experiments on edges2shoes.
Release Fast GAN Compression. Include the GauGAN experiments on COCO-stuff.
Simplify the lite pipeline (GAN Compression Lite)! Release new models of GAN Compression Lite.
Release the simplified pipeline for the cycleGAN and pix2pix model.
Support GauGAN training.
Fix a bug in weight_transfer.py
.
The official Pytorch implementation of GAN Compression v1.0.
Compared to the previous version, we corrected metric naming (mAP to mIoU), and updated Cityscapes mIoU valuation protocol (DRN(upsample(G(x))) -> upsample(DRN(G(x)))
). We also added an option for the searching which supports only evaluating the sub-networks under a certain budget within the "once-for-all" network.