Gram matrix calculation

[AlexTS1980]

I have a strange problem with Gram matrix and style loss. If I define the Gram matrix for style and generated images as :

def gram(tensor):
    B, C, H, W = tensor.shape
    x = tensor.view(B*C, H*W)
    return 1/(2*B*C*H*W) * torch.mm(x, x.t())

B=1, C=number of feature maps

style loss very quickly converges and the resulting output is a white noise essentially OR style loss is very small to begin with and the results is the same. Only when I remove C in the denominator, style loss keeps changing throught the stylization training process. What did I do wrong? The style weight=10000, content weight=5, Adam learning rate=1. I tried slightly different parameters, but the problem remains.

[Rusty Mina]

Generally, the resulting (gram matrix) x transpose(gram matrix) should be of shape C by C. Nonetheless, since B = 1, you may ignore that issue.

Unto to the problem, I think you'll need to use much much bigger style weights (~ 100 million).

But then again, easiest solution would be drop the 1/(2*B*C*H*W) and use smaller values of STYLE_WEIGHT.

[AlexTS1980]

No it turned out I incorrectly loaded the weights.