_loss.py

import torch


class CustomLoss(object):

    #### user can custom weight ####

    def __init__(self, w1, w2):

        self.w1 = w1
        self.w2 = w2
        self.eps = 1e-7

    def custom_loss_1(self, output, target):

        ### baseline is the first value of target variable ###

        outsgn = torch.sign(output - target[0] * output.shape[1])
        trgsgn = torch.sign(target - target[0] * output.shape[1])
        binary = outsgn != trgsgn

        w1 = torch.ones(*binary.shape) * self.w1
        w2 = torch.ones(*binary.shape) * self.w2
        w1 = w1.to("cuda")
        w2 = w2.to("cuda")

        binary = torch.where(binary < self.eps, w1, w2)
        se = (output - target) ** 2
        loss = torch.mean(se * binary)

        return loss

    def custom_loss_2(self, output, target):

        ### baseline is the previous value of target variable ###

        zero = torch.zeros(1, output.shape[1]).to("cuda")
        binary = torch.sign(output[1:] - target[:-1]) != torch.sign(
            target[1:] - target[:-1]
        )
        binary = torch.cat((zero, binary))
        w1 = torch.ones(*binary.shape) * self.w1
        w2 = torch.ones(*binary.shape) * self.w2
        w1 = w1.to("cuda")
        w2 = w2.to("cuda")

        binary = torch.where(binary < self.eps, w1, w2)
        se = (output - target) ** 2
        loss = torch.mean(se * binary)

        return loss