Features of latest update

[lucidrains]

Features of latest update

[dave7895]

oh, that looks nice. Thank you for the insights.

[lucidrains]

@dave7895 this feature is actually a hack that people in the GAN field use to make the generated images look better. what they do is they estimate the average latent vector, and make sure that the random latent vector you sample does not stray too far from the average. --trunc-psi ranges from 0 to 1, and denotes that distance.

[dave7895]

I am not really experienced with the whole field, rather just starting out, so I do not really know what a latent vector is. Would just a very general short description or explanation be possible if it is not too much?

[lucidrains]

@dave7895 the generator actually learns to create the images from a bag of numbers (512 to be exact). we call that collection of numbers the latent vector.

latent vector -> generator -> image
image -> discriminator -> 1 or 0, (real or fake)

[dave7895]

These numbers come from the nodes of the neural net?

[lucidrains]

@dave7895 you pass this latent vector into the net, you get to choose them!

[dave7895]

Is it so that the network chooses one of the 512 and with psi-trunc you make the space of choosing for centered and try to remove outliers?

[lucidrains]

@dave7895 yes, actually I randomly sample 512 numbers (z), and the outliers are removed once they are transformer to w (style vector)

[lucidrains]

@dave7895 each individual image actually comes from a specific set of 512 numbers. i choose them randomly, but you can set those numbers yourself

[dave7895]

So it would be possible to use the exact same set of numbers on different models and then watch the about the same face get generated by different generations?

[lucidrains]

@dave7895 we don't have much control over how the generator decides to layout the latent space, so every time you train a new network, it may decide to put the same face at a different coordinate

[dave7895]

That I can understand. Just like brains being different between siblings for example. But how is it with the same network a thousand iterations later?

[lucidrains]

@dave7895 ohh, what do you mean? do you mean how is the face for the same latent vector of the same network trained for 1k iterations?

[dave7895]

The actual question was if the latent space, which defines the possible latent vector due to its dimensions if I understood correctly, changes during training. If it does not change then you could feed the same numbers each time at least once, and you could see one of the faces evolving as the network evolves.

[lucidrains]

@dave7895 oh yes, many people do that with GANs already! https://pythonawesome.com/content/images/2019/08/celeba_dragan.gif