xuqiantong/gan-metrics

An empirical study on evaluation metrics of generative adversarial networks.

gan-metrics's Issues

dataset

Is it suitable for all pictures? Tahnk you.

Incomplete Inception_v3 architecture..

I was going through your feature calculation for Inception_v3 in metrics.py and found out that you did not use the 'inception.Mixed_6e' block in the feature calculation. Can you comment on the reasons you chose to do that?

What does the function "distance" really measure.

I'm wondring what is the real meaning of the function distance. at first I thought it was an Euclidean distance between two features spaces but when reading the code it is not . I also don't get it why Mxx != 0 ! the distance to itself should be always 0 otherwise its only a function name but not a real distance.

I hope somone could clarify this concept. I will be glad to have any documentation, paper or articles
Thnx

Out of memory when "--sampleSize 10000"

The TF version of GAN metrics, e.g. IS, FID, are using 10000 samples, while sampleSize==2000 in the demo.
The setting of "sampleSize==10000" may cause out of memory problem, as dozens of GB mem is requested by "metrics.py".
Could the code used to reproduce the "sampleSize==10000" experiments as TF version does?

Thanks for your great work.

your paper's collation map

Hello, how did you get the fractional polyline diagram of pixel space and convolution space in your paper？How should I operate to get the same result as you?
How do you operate on changes in the number of false images contained in the overall dataset? I didn't find the relevant code?
Thank you very much for your help.

Trying to run the code in dcgan_demo.py but the samples are not created in the 'real' folder of training set.

Apparently no real sample is generated in the target training folder when the code is running in training sessions. Please check it and possibly fix it.

Output of Inception score is different that the torch inception model

Hi,

In metrics.py where you have defined the inception feature, inception.Mixed_6e is missing and this result is getting different outputs which is not comparable to other methods.

FID should compute in conv space instead of softmax space

I noticed the line 428 of metric.py which compute FID in softmax space, the original FID should compute in conv space instead of softmax space, and I find no description about this.