Style loss: MSE or SSE

[Mona77]

Hi, I was wondering why you have calculated sum of squared in style loss calculations:

m_loss = sse(d_mean, s_mean) / batch_size # normalized w.r.t. batch size

While in torch implementation MSE is being used, this line:

self.mean_criterion = nn.MSECriterion() self.mean_loss = self.mean_criterion:forward(self.input_mean, self.target_mean)

[eridgd]

It's been too long for me to remember exactly why I implemented things the way I did, but my guess is it's a mistake. The results do still seem to be comparable to the Torch version, but at some point I'll retrain with MSE and see how it affects the output. Good catch!

[Mona77]

@eridgd, I confirm that. I trained your network as it is, and the results are very similar to the Torch version.
Please let me know. Thank you!