Indices of layers

[JanineCHEN]

Thank you for this great work.

I have noticed that when doing manipulation, for --layerwise_manipulation, you indicate a specific layer using the index of the layer to perform manipulation such as 6-11 in the example. i am just wondering how did you retrieve the correct index of a particular layer, assuming using a different generator model?

Any insights would be greatly appreciated!

[ShenYujun]

We use the proposed re-scoring technique to conduct a layer-wise analysis. In this way, the layers with the highest response are treated as the most relevant layers.

[JanineCHEN]

We use the proposed re-scoring technique to conduct a layer-wise analysis. In this way, the layers with the highest response are treated as the most relevant layers.

Thank you @ShenYujun for your kind explanation. I am just wondering how did you get the list of layer indices in the first place. Also I am not quite clear if 6-11 indicates layer 6 to layer 11, or the sublayer 11 of layer 6?

By the way, I have trained my own StyleGAN model using the official PyTorch version StyleGAN provided here by facebook search. Could you kindly advise me if I can embed my trained model into your pipeline? I am not sure if their model and yours follow the same mechanism?

[ShenYujun]

The first convolution layer used in the generator is Layer 0, the second is Layer 1, and so on.

I am not sure whether it is the "official" PyTorch implementation of StyleGAN. As far as I know, the authors of StyleGAN only release one official codebase, which is using TensorFlow. Our model exactly aligns with theirs and we also support converting tf weights to pth weights in our codebase ;)

[JanineCHEN]

The first convolution layer used in the generator is Layer 0, the second is Layer 1, and so on.

I am not sure whether it is the "official" PyTorch implementation of StyleGAN. As far as I know, the authors of StyleGAN only release one official codebase, which is using TensorFlow. Our model exactly aligns with theirs and we also support converting tf weights to pth weights in our codebase ;)

Hi @ShenYujun , thanks a lot for your instant reply. I have noticed that you also implemented biggan in your pipeline, but seems like you haven't included biggan in this repo? Actually, I opened another issue inquiring this question. I will close this issue then. And it would also be great if you could kindly shift to that issue and share some more information on the biggan implementation.

Thank you!