Domain Transfer Networks (DTN)

• UNSUPERVISED CROSS-DOMAIN IMAGE GENERATION, ICLR 2017

  Install

• install pytorch and pytorch.vision

Run

• Train function f in source domain(SVHN), first

  • CUDA_VISIBLE_DEVICES=x python main_recog.py --dataset svhn --dataroot /path/to/svhn/extra/ --valDataroot /path/to/svhn/test/ --exp recog_svhn
  • Datasets will be downloaded automatically in the paths you specified
  • Resulting model is saved in recog_svhn directory named like netE_epoch_xx.pth
  • You will reach at ~95.xx % of accuracy.

• And then, train DTN

  • CUDA_VISIBLE_DEVICES=x python main_dtnetgan.py --datasetA svhn --datarootA /path/to/svhn/extra/ --valDatarootA /path/to/svhn/test/ --datasetB mnist --datarootB /path/to/mnist/train/ --valDatarootB /path/to/test/ --netE /path/to/previously/trained/model/netE_epoch_xx.pth --exp S2M --crossentropy

Results (Domain transfer)

• Randomly selected samples in source domain

• Domain transferred samples from corresponding inputs

Results (Style transfer)

• Style transfer as a specific case of the domain transfer
• not yet implemented

NOTE

• We used crossentropy loss computing L_CONST (i.e. Eq.5 in the paper)
• This idea was borrowed from Plug-and-Play Generative Network
• You can easily change direction of domain transfer such that MNIST to SVHN
• CUDA_VISIBLE_DEVICES=x python main_recog.py --dataset mnist --dataroot /path/to/mnist/train/ --valDataroot /path/to/mnist/test/ --exp recog_mnist
• CUDA_VISIBLE_DEVICES=x python main_dtnetgan.py --datasetA mnist --datarootA /path/to/mnist/train/ --valDatarootA /path/to/mnist/test/ --datasetB svhn --datarootB /path/to/svhn/extra/ --valDatarootB /path/to/svhn/test/ --netE /path/to/pretrained/model/netE_epoch_xx.pth --exp M2S

Reference

• dcgan.pytorch
• imagenet.pytorch
• FANTASTIC pytorch pytorch doc
• genhacks from soumith
• Plug and Play Generative Network