search

ProGamerGov / neural-style-pt

PyTorch implementation of neural style transfer algorithm
MIT License
833 stars 178 forks source link

normalize_weights #55

Closed spot92 closed 4 years ago

spot92 commented 4 years ago

Does normalize_weights work the same as normalize_gradients from jcjohnson? I have seen you posting about setting content_weight to 0 in order to achieve the same effect as normalize_gradients, though. So if this is not the case, what does normalize_weights do?

ProGamerGov commented 4 years ago

It's not the same thing as -normalize_gradients and thus far no one has been able to get that feature working in PyTorch.

By default, the same style and content weight values that you choose are used for all the content and style layers. The -normalize_weights parameter alters weight values on a per layer basis, by dividing the weight for each layer by the number of channels in the layer. Lower layers have fewer channels and higher layer have more channels.

The feature was inspired by what Leon Gatys does in: NeuralImageSynthesis

ProGamerGov commented 4 years ago

https://github.com/ProGamerGov/neural-style-pt/issues/26#issuecomment-538312130

https://github.com/ProGamerGov/neural-style-pt/issues/43

ProGamerGov commented 3 years ago

@spot92 I finally figure out how to recreate the -normalize_gradients parameter in PyTorch, and I've added it to neural-style-pt!