How to implement custom loss function

1

Hello,I am using a Unet network and need to write a loss function myself, this loss function needs to involve calculations that convert to numpy.

Summary 
def my_mse_loss(input, target):
    a = input.data[0].numpy().flatten()
    b = target.data[0].numpy().flatten()
    loss = 0
    for i in range(len(a)):
        x = a[i]-b[i]
        loss += abs(x**2)
    return torch.Tensor(loss) / input.data.nelement()

Or can you write an example for me?More detailed.

2

Sorry but why do you need the numpy operation
I think you can just use toroch.nn.MSELoss
http://pytorch.org/docs/0.3.0/nn.html?highlight=mse#torch.nn.MSELoss

Or do I misunderstand this topic?

3

I want to do other operations, here is just an example

4

What kind of operations do you need from numpy that are not available in Pytorch?
It would make the implementation easier, if it’s possible to use pure Pytorch.