search

paarthneekhara / text-to-image

Text to image synthesis using thought vectors
MIT License
2.16k stars 394 forks source link

Input of Discriminator #4

Closed chingyaoc closed 8 years ago

chingyaoc commented 8 years ago

Hi, here's another question. Here the discriminator's input is wrong image and right text. However according to the original paper (page 5), they use real image and wrong text which is different from your implementation.

paarthneekhara commented 8 years ago

Hi, The discriminator has to learn to identify real images with the right text (output should be 1). The output should be 0 if the text does not match with the image, or if the image is fake (generated). Right Image with wrong text is equivalent to wrong image with right text (i.e image and text do not correspond to each other). Wrong image and right text, are used in the implementation of the paper as well (https://github.com/reedscot/icml2016). It just an implmentation convinience.

chingyaoc commented 8 years ago

Thanks for your reply! Really helps me a lot.