conv layer used for computing the content loss about chainer-fast-neuralstyle HOT 3 OPEN

Fig.2. in the paper shows feature map of layer relu3_3 is used.
According to Fig.3., as reconstruction from relu2_2 looks almost the same as the input image, I think relu2_2 needs more abstraction for the "feature". I think either relu3_3 or relu4_3 is good.

from chainer-fast-neuralstyle.

Thanks! I experimented on relu3_3 and relu2_2, and yes, you are right, relu2_2 cannot represent the content very well.

Have you solved the batch size issue? Do you think it would be relevant to the batch_normalization function in your transform net?

from chainer-fast-neuralstyle.

@littlekobe are you saying you found a way to get better feature extraction in this implementation using relu3_3? If so, how do I activate that layer in the model or generation? Whichever it is, I'll do what it takes to get that van gogh starry night to be a little more accurate, even if training or generating takes a little longer.

from chainer-fast-neuralstyle.