Latent Mapper never trains

[ਰਾਹੁਲ ਭੇਲੇ]

Hi, thanks for StyleCLIP work.

But when I train Latent Mapper using instructions under Editing via Latent Mapper section my latent mapper just never trains.

I get a best_model.pt of 145.34 MB. whereas latent mapper (including all 3 spatial control levels) is just 12.6 MB. So I put extract the mapper parameters and load them in a new latent_mappers.LevelsMapper instance. But then I do inference while traversing in latent space in the direction of what new LevelsMapper well trained. Something like this: new_latent = latent_code_init + mapper(latent_code_init) * I but there are not manipulations happening in image.

Here are samples (initial image and manipulated image).

1) For description "bob-cut hairstyle"

2) For description "smiling face"

[ਰਾਹੁਲ ਭੇਲੇ]

Just found out that it does work but only if I load decoder parameters too. But shouldn't that be locked during training? I mean, aren't we supposed to change only latent mapper to manipulate different aspect of an image?

[stas-polukeev]

hi! could you provide the commands you used for training, please? I'm trying the same, but the generator doesn't seem to work correctly