Hi, did anyone successfully implemented the pixel-wise weighted loss function that was described in the Unet paper?
Thanks
Hi,
Yes of course.
Here is some link to implementation of Github community. They have trained their model and used, so they have to be correct.
link1: https://github.com/milesial/Pytorch-UNet
link2: https://github.com/usuyama/pytorch-unet
link3: https://github.com/jvanvugt/pytorch-unet
link4: https://www.kaggle.com/witwitchayakarn/u-net-with-pytorch
If you did not understand any part or you have other questions, feel free to ask.
Hi Nikronic, Thanks for the links!
However, None of these Unet implementation are using the pixel-weighted soft-max cross-entropy loss that is defined in the Unet paper (page 5).
I’ve tried to implement it myself using a modified version of this code to compute the weights which I multiply by the CrossEntropyLoss:
loss = nn.CrossEntropyLoss(reduction='none')(output, target)
loss = torch.mean(loss * weights)
where the output is the output tensor from the Unet.
When I am using this implementation my model is a complete failure (F1 score = 0).
So, I’ve replace the CrossEntropyLoss by:
m = nn.Sigmoid()
loss_per_pixel = torch.nn.BCELoss(reduce=False)(m(input), target) * weights
In this case my model converge to a relatively good result but this result is less accurate than if I’ve used a non-weighted loss function.
So, I am wondering if I am doing something wrong.
Thanks for your help!
I’ve implemented a weighted pixelwise NLLLoss a while ago in this topic. The code is quite old (e.g. don’t use Variables
anymore), but might be a good starter for your custom loss.
Hey,
this is an old thread but I believe Sam’s formulation is right:
I’ve used this loss myself in the past and it worked out fairly well, assuming the scale of the weights is
“correct”. Make sure to pair it with something like a dice loss so that the CE can be used to incentivize precision around the edges.