FID for GAN trained on MNIST

[Nagabhushan S N]

Hi,
I'm training a DC GAN on MNIST dataset. I want to compute FID for my model. In this case should I use the Inception network itself or should I use a different classifier trained on MNIST dataset?

[Martin Heusel]

Hi,

the Inception model expects a 3-channel image. One solution is to triple the MNIST single channel and motify the generator to output in the same 3-channel space as well but then you're not generating the original MNIST images anymore. So yes, you better train a different classifier on the original data and use the statistics of one of the last layers, or train an autoencoder and use the bottleneck activations. The resulting FID however, depends on this specific classifier/autoencoder and is not really comparable with FIDs from other classifiers/autoencoders.

[Nagabhushan S N]

Hi, Thank you for the clarification. Can you explain what is the intuition for selecting bottleneck activations from an autoencoder?

[Martin Heusel]

The bottleneck activiations are lower dimensional representations of the input data. The hope is, that the bottleneck encodes most of the information of the input data in a lower dim and is therefore a good candidate for the FID statistics (mean, variance).

[Nagabhushan S N]

Oh! yeah, that's cool. I'll try that. Thanks :)
Closing the issue