How to define loss

User-defined loss can be anything derived from torch.nn.Module with defined forward method, which takes the model output prediction and label target as input and returns the loss value. The loss value should be a scalar tensor. For example, we can define the mean absolute scaled error (MASE) loss as follows:

import torch
import torch.nn as nn

class MASELoss(nn.Module):
    '''Mean Absolute Scaled Error Loss'''
    def __init__(self, min_number=1e-8):
        super(MASELoss, self).__init__()
        self.min_number = min_number # floor for denominator to prevent inf losses

    def forward(self, prediction, target):
        numerator = torch.mean( torch.abs(prediction-target) )
        denominator = torch.mean( torch.abs(torch.diff(target),n=1) )
        denominator = torch.maximum(denominator,torch.mul(torch.ones_like(denominator),self.min_number))
        return torch.divide(numerator,denominator)

To use the loss function during the training, you need to provide the absolute/relative path to the loss definition file and the name of the loss class. For example, to use the MASE loss defined above, you can add the following lines to the server configuration file:

client_configs:
    train_configs:
        ...
        # Loss function
        loss_fn_path: "<path_to_mase_loss>.py"
        loss_fn_name: "MASELoss"
        loss_fn_kwargs:
            min_number: 1e-8
    ...