search

huangzehao / caffe-vdsr

A Caffe-based implementation of very deep convolution network for image super-resolution
MIT License
273 stars 134 forks source link

loss during training #18

Closed xjtuljy closed 7 years ago

xjtuljy commented 7 years ago

I am curious about the loss during training: what's the typical loss when it converges in your work (I think you use 256X256 size image for training?)?

Many thanks.

Jianyu

huangzehao commented 7 years ago

x2 loss: 0.16~0.17 training sample size 41x41

xjtuljy commented 7 years ago

Thanks @huangzehao . What's the gray level of your input image? is it [0,1] or [0,255]? if my input is between [0,1], since the EuclideanLoss E = 1/2N×sum((y-y')^2), shall I expect the max loss to be 0.5?

Jianyu

huangzehao commented 7 years ago

Hi, the gray level of input image is [0,1]. The max loss is not 0.5, since the output of the network is not limit to [0,1].

xjtuljy commented 7 years ago

Thanks. I am using similar networks, with EuclideanLoss layer. My input is [0,255], but I stretch that into [0,1], using
transform_param { scale: 0.00390625 } In that way, I end up with loss ~1k when convergence, with fixed learning rate, without fine-tuning. The output during testing still seems to make sense. So I wonder is this 1k loss I got the result of 256*Euclidean loss?