Page MenuHomec4science
Files F90712396

Loss.py
No OneTemporary
Actions

File Metadata

Created
Mon, Nov 4, 02:22

Loss.py
View Options

# -*- coding: utf-8 -*-
"""
@author: srpv
contact: vigneashwara.solairajapandiyan@empa.ch, vigneashpandiyan@gmail.com
The codes in this following script will be used for the publication of the following work
"Qualify-As-You-Go: Sensor Fusion of Optical and Acoustic Signatures with Contrastive Deep Learning for Multi-Material Composition Monitoring in Laser Powder Bed Fusion Process"
@any reuse of this code should be authorized by the first owner, code author
"""
#libraries to import
import torch.nn as nn
import torch
from torch.nn import functional as F
from typing import Tuple
from torch import nn, Tensor
def convert_label_to_similarity(normed_feature: Tensor, label: Tensor) -> Tuple[Tensor, Tensor]:
"""
Converts the label tensor into similarity matrices for positive and negative pairs.
Args:
normed_feature (Tensor): The normalized feature tensor.
label (Tensor): The label tensor.
Returns:
Tuple[Tensor, Tensor]: A tuple containing the similarity matrices for positive and negative pairs.
"""
# Calculate the similarity matrix
similarity_matrix = normed_feature @ normed_feature.transpose(1, 0)
# Create a label matrix
label_matrix = label.unsqueeze(1) == label.unsqueeze(0)
# Create positive and negative matrices
positive_matrix = label_matrix.triu(diagonal=1)
negative_matrix = label_matrix.logical_not().triu(diagonal=1)
# Flatten the similarity and label matrices
similarity_matrix = similarity_matrix.view(-1)
positive_matrix = positive_matrix.view(-1)
negative_matrix = negative_matrix.view(-1)
# Return the similarity matrices for positive and negative pairs
return similarity_matrix[positive_matrix], similarity_matrix[negative_matrix]
class CircleLoss(nn.Module):
"""
CircleLoss is a custom loss function used for contrastive learning.
Args:
m (float): Margin parameter for the loss function.
gamma (float): Scaling parameter for the loss function.
Inputs:
sp (Tensor): Positive similarity tensor.
sn (Tensor): Negative similarity tensor.
Returns:
Tensor: The computed loss value.
"""
def __init__(self, m: float, gamma: float) -> None:
super(CircleLoss, self).__init__()
self.m = m
self.gamma = gamma
self.soft_plus = nn.Softplus()
def forward(self, sp: Tensor, sn: Tensor) -> Tensor:
ap = torch.clamp_min(- sp.detach() + 1 + self.m, min=0.)
an = torch.clamp_min(sn.detach() + self.m, min=0.)
delta_p = 1 - self.m
delta_n = self.m
logit_p = - ap * (sp - delta_p) * self.gamma
logit_n = an * (sn - delta_n) * self.gamma
loss = self.soft_plus(torch.logsumexp(logit_n, dim=0) + torch.logsumexp(logit_p, dim=0))
return loss

Event Timeline

c4science · Help