PConv-Keras
PConv-Keras copied to clipboard
MathiasGruber
the range of loss when training
For my original implementation, the input image and target image are normalized to be in the range of [-1,1]; for the input for VGG based loss, it is also in that range. Thus, the loss won't be too big. For my case, it is between [0, 3].
@liuguilin1225 In the paper, you used a VGG model trained on pytorch which requires the image pre-processing you point to on your website, namely this:
- dividing by 255,
- subtracting a per-channel mean,
- dividing by a per-channel std.
It does not seem to lead to the [-1,1] range but rather to [-2.12, 2.25], [-2.04, 2.43] and [-1.80, 2.64] for the R, G and B channels. Am I missing something or did you use another pre-processing?
@ezavarygin sorry for the late reply. I mainly followed the link you posted, I guess that is how the original VGG model was trained. https://github.com/pytorch/examples/blob/master/imagenet/main.py#L198
It doesn't necessarily to be in [-1, 1]. Especially for computing perceptual loss and style loss, we just need to make sure the two compared images are using the same preprocessing scheme.
I've ported the VGG16 model from pytorch to keras now, and also updated the PConv2d layer to use local normalization. Loss ranges should be significantly different now, eh?
I'm facing the same problem while training from scratch the loss starts at [10,9] the model gives brown images for the initial epoch. So I started finetuning with the provided imagenet weights but the loss still seems to start at [10,9] is this expected behaviour ? Also is the hyperparams of the fixmyphoto.ai same as in the notebook, because i'm getting good results with the website, but with the model I'm not getting similar results.
