loss.py

"""
Losses
"""
# pylint: disable=C0301,C0103,R0902,R0915,W0221,W0622


##
# LIBRARIES
import torch

##
def l1_loss(input, target):
    """ L1 Loss without reduce flag.

    Args:
        input (FloatTensor): Input tensor
        target (FloatTensor): Output tensor

    Returns:
        [FloatTensor]: L1 distance between input and output
    """

    return torch.mean(torch.abs(input - target))

##
def l2_loss(input, target, size_average=True):
    """ L2 Loss without reduce flag.

    Args:
        input (FloatTensor): Input tensor
        target (FloatTensor): Output tensor

    Returns:
        [FloatTensor]: L2 distance between input and output
    """
    if size_average:
        return torch.mean(torch.pow((input-target), 2))
    else:
        return torch.pow((input-target), 2)

def loss_func(adj, A_hat, attrs, X_hat):
    # Attribute reconstruction loss
    diff_attribute = torch.pow(X_hat - attrs, 2)
    attribute_reconstruction_errors = torch.sqrt(torch.sum(diff_attribute, 1))
    attribute_cost = torch.mean(attribute_reconstruction_errors)

    # structure reconstruction loss
    diff_structure = torch.pow(A_hat - adj, 2)
    structure_reconstruction_errors = torch.sqrt(torch.sum(diff_structure, 1))
    structure_cost = torch.mean(structure_reconstruction_errors)


    return structure_cost, attribute_cost


def loss_cal(x, x_aug):
    
    T = 0.2
    batch_size, _ = x.size()
    x_abs = x.norm(dim=1)
    x_aug_abs = x_aug.norm(dim=1)
    sim_matrix = torch.einsum('ik,jk->ij', x, x_aug) / torch.einsum('i,j->ij', x_abs, x_aug_abs)
    sim_matrix = torch.exp(sim_matrix / T)
    pos_sim = sim_matrix[range(batch_size), range(batch_size)]
    loss = pos_sim / (sim_matrix.sum(dim=1) - pos_sim)
    loss = - torch.log(loss).mean()
    return loss