I am trying to use feature reconstruction and style reconstruction losses on my model. For this, I followed the example code on PyTorch website for “Neural Style Transfer”.
Although the feature loss is calculated without problem, the style loss is always zero. And I could not find the reason since everything looks okay in the implementation. The calculation formulas are the same with the proposed mathematical formulas. Besides, as you know, the style and feature losses are almost same in terms of calculation except Gram matrix step in style loss and no problem in feature loss.
Could anyone help me about this situation?