search

f90 / FactorGAN

FactorGAN - Training GANs with missing data
MIT License
35 stars 3 forks source link

How can i improve my results? #5

Open kalai2033 opened 4 years ago

kalai2033 commented 4 years ago

i have a custom dataset with 300 image pairs. I have used to pix2pix/cycleGAN models for image translation task. I have got some good interesting results. But the generation of image pairs is very time consuming and costly. When I have read your paper it seems your model does image to image translation in a semi-supervised manner with less data points (image pairs). I trained a model with 300 image pairs with --factorGAN flag as 1 and --num_joint_samples flag as 300. But my results were not good compared to pix2pix and cycleGAN. Can you give suggestion to improve my results?

f90 commented 4 years ago

Are you using the same generator model in cycleGAN and FactorGAN? The default neural network used in FactorGAN might not be particularly strong compared to the ones used in the original cycleGAN or pix2pix papers. That has a big influence on performance. Similarly, the discriminator architecture should be made somewhat comparable.