torch_tcce.py

#!/usr/bin/env python3
 
 
 
import torch
import torch.nn as nn
import numpy as np
 
 
def tcc(y: torch.tensor, y_true: torch.LongTensor, n: int) -> torch.tensor:
    """
    Calculates loss using the required number of Taylor terms of cross entropy loss.
 
    Parameters:
        y(torch.tensor): The probabilities predicted by ML model.
        y_true(torch.LongTensor): The truth values provided for training purposes (1D tensor).
        n(int): Number of terms to to be taken for the Taylor Series.
 
    Returns:
        A torch tesor with the value of the calculated Taylor cross entropy loss.
 
    Note:
        With n = 0, this returns the regular cross entropy loss.
 
    """
    loss = torch.zeros(len(y_true))
    if torch.cuda.is_available():
        loss = loss.to("cuda")
    ProbTrue = y[np.arange(len(y_true)), y_true]
    if n != 0:
        for i in range(1, n + 1):
            loss += torch.pow(1 - ProbTrue, i) / i
    elif n == 0:
        loss = -1 * torch.log(ProbTrue)
    loss = torch.sum(loss)
    return loss / len(y)
 
 
class TCCE(nn.Module):
    """
    Class for calculation of Taylor cross entropy loss.
 
    Attributes:
        n (int): Number of Taylor series terms to be used for loss calculation.
 
    """
 
    def __init__(self, n: int = 0):
        """
        Initialize the loss class.
 
        Parameters:
            n (int)(optional): Number of Taylor series terms to be used for loss calculation.
 
        """
        super().__init__()
        
        self.n = n
 
    def forward(self, y: torch.tensor, y_true: torch.LongTensor) -> torch.tensor:
        """
        Calculates the Taylor categorical cross entropy loss.
 
        Parameters:
            y(torch.tensor): Tensor containing the output of the model.
            y_true(torch.tensor): 1D tensor containing the truth value for a given set of features.
 
        Returns:
            The calculated loss as a torch tensor.
        """
        return tcc(y, y_true, self.n)