dmitryulyanov / age Goto Github PK

Hi,

Just for curiosity, I wonder if I get the right pipeline to check the inception score on CelebA dataset as you presented in the paper.

Should I use the Inception model that is pretrained on ImageNet or trained on CelebA from scratch by myself?

Thanks in advance :)

KLD appears to use variance in place of standard deviation. utils.var() computes variance as squared distance from mean. Then it's squared again in the KLN01Loss module. Should it be (in the default 'qp' direction):

t1 = samples_var + samples_mean.pow(2)
t2 = -samples_var.log()

KL = (t1 + t2 - 1).mean()/2

?

(Additionally, the paper gives the KLD as a sum but here it's a mean, changing the meaning of the hyperparameters weighting the reconstruction losses)

I was not able to get the CIFAR10 example to run with the defaults, since the expand_as call expects the expanding dimension to be greater or equal to the first dimension (and the defaults have the batch size as 64 and nz as 128). If I set the batch size and the nz sizes the same, it seems to work fine.

Hi Ulyanov, today I just try to download the pre-trained weight. But there is some issue

Hi Dmitry,

I have one question. Why using ((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) for the Normalize in the SVHN?

That values are the ones you want your input to be modified so the fit into that distribution? Or you need to imput your actual mean and standard deviation of each channel to get normilized?

Also, why sometimes there are other configuration of values?
I don't understand why still with those parameters I have negative values in the images.

Thanks a lot in advance.

Pablo

Hi! nice work! Are you releasing training code?

why some of my reconstructions are just blank images? Using a grayscale dataset with default parameters