I had trained a styleGAN in my dataset, but in this ckpt, we miss the key named 'latent_avg', how to get it?

[omertov]

I had trained a styleGAN in my dataset, but in this ckpt, we miss the key named 'latent_avg', how to get it?

[yuval-alaluf]

To add to @omertov's answer, when you initialize the network in the following lines:

encoder4editing/training/coach.py

Lines 37 to 38 in 73c60bd

	# Initialize network
	self.net = pSp(self.opts).to(self.device)

you can add the following snippet:

if self.net.latent_avg is None:			
    self.net.latent_avg = self.net.decoder.mean_latent(int(1e5))[0].detach()

This will estimate the average latent and store it as part of the network. Then training can continue as is normal.

[omertov]

Hi @ssxxx1a.
The average latent is obtained from sampling many (usually thousands) of latent codes and averaging them to get the average code.
Usually the repositories provide a build in function for obtaining this average code, for example see this function in rosinality's repository.

After obtaining the average code, you should save it as a part of the checkpoint under the 'latent_avg' key.

Best,
Omer

[ssxxx1a]

ohhh!
thanks both for your reply !