search

ceshine / fast-neural-style

Pytorch Implementation of Perceptual Losses for Real-Time Style Transfer and Super-Resolution
162 stars 34 forks source link

Why not divide the size ? #4

Closed FantasyJXF closed 5 years ago

FantasyJXF commented 5 years ago

In the original paper and the most stared TF implementation, the content loss and style loss should divide the feature map size (C,H,W).

Why don't you do this?

ceshine commented 5 years ago

I used STYLE_WEIGHTS to set weights for loss from different layers. You can incorporate the size of the feature maps into those layers to achieve the same effect.

In this way, the layer weights become part of hyper-parameters and can be picked empirically.

FantasyJXF commented 5 years ago

That's smart, use STYLE_WIGHTS of different style layers, it could get the similar result to a certain degree.

And you add tv loss, though not the standard format.

Only one thing, the transferred style image in your notebook didn't look so good, do you have any pretrained model which looks nice?

ceshine commented 5 years ago

I don't think I still have any pretrained models in my computer. There are already a lot of newer and better style transfer algorithms around. Here's one that I've re-implemented in PyTorch: PyTorch Implementation of Style Transfer as Optimal Transport.

FantasyJXF commented 5 years ago

That's very nice!

FantasyJXF commented 5 years ago

The Wasserstein distance model acted like the original style transfer, it's a iterated update procedure, which takes much time?

ceshine commented 5 years ago

It's just another loss function. See this notebook for explanation.

FantasyJXF commented 5 years ago

I use Tesla P40(24G graphic memory) to inference the style image, it's unbelievable that the model went out "RUN OUT OF MEMORY" error

ceshine commented 5 years ago

That didn't happen in my case. Maybe the content image you use was too big?