2021-06-04 22:15:57 +02:00

95 lines
3.1 KiB
Python

"""ProtoTorch loss functions."""
import torch
def _get_matcher(targets, labels):
"""Returns a boolean tensor."""
matcher = torch.eq(targets.unsqueeze(dim=1), labels)
if labels.ndim == 2:
# if the labels are one-hot vectors
num_classes = targets.size()[1]
matcher = torch.eq(torch.sum(matcher, dim=-1), num_classes)
return matcher
def _get_dp_dm(distances, targets, plabels, with_indices=False):
"""Returns the d+ and d- values for a batch of distances."""
matcher = _get_matcher(targets, plabels)
not_matcher = torch.bitwise_not(matcher)
inf = torch.full_like(distances, fill_value=float("inf"))
d_matching = torch.where(matcher, distances, inf)
d_unmatching = torch.where(not_matcher, distances, inf)
dp = torch.min(d_matching, dim=-1, keepdim=True)
dm = torch.min(d_unmatching, dim=-1, keepdim=True)
if with_indices:
return dp, dm
return dp.values, dm.values
def glvq_loss(distances, target_labels, prototype_labels):
"""GLVQ loss function with support for one-hot labels."""
dp, dm = _get_dp_dm(distances, target_labels, prototype_labels)
mu = (dp - dm) / (dp + dm)
return mu
def lvq1_loss(distances, target_labels, prototype_labels):
"""LVQ1 loss function with support for one-hot labels.
See Section 4 [Sado&Yamada]
https://papers.nips.cc/paper/1995/file/9c3b1830513cc3b8fc4b76635d32e692-Paper.pdf
"""
dp, dm = _get_dp_dm(distances, target_labels, prototype_labels)
mu = dp
mu[dp > dm] = -dm[dp > dm]
return mu
def lvq21_loss(distances, target_labels, prototype_labels):
"""LVQ2.1 loss function with support for one-hot labels.
See Section 4 [Sado&Yamada]
https://papers.nips.cc/paper/1995/file/9c3b1830513cc3b8fc4b76635d32e692-Paper.pdf
"""
dp, dm = _get_dp_dm(distances, target_labels, prototype_labels)
mu = dp - dm
return mu
# Probabilistic
def _get_class_probabilities(probabilities, targets, prototype_labels):
# Create Label Mapping
uniques = prototype_labels.unique(sorted=True).tolist()
key_val = {key: val for key, val in zip(uniques, range(len(uniques)))}
target_indices = torch.LongTensor(list(map(key_val.get, targets.tolist())))
whole = probabilities.sum(dim=1)
correct = probabilities[torch.arange(len(probabilities)), target_indices]
wrong = whole - correct
return whole, correct, wrong
def nllr_loss(probabilities, targets, prototype_labels):
"""Compute the Negative Log-Likelihood Ratio loss."""
_, correct, wrong = _get_class_probabilities(probabilities, targets,
prototype_labels)
likelihood = correct / wrong
log_likelihood = torch.log(likelihood)
return -1.0 * log_likelihood
def rslvq_loss(probabilities, targets, prototype_labels):
"""Compute the Robust Soft Learning Vector Quantization (RSLVQ) loss."""
whole, correct, _ = _get_class_probabilities(probabilities, targets,
prototype_labels)
likelihood = correct / whole
log_likelihood = torch.log(likelihood)
return -1.0 * log_likelihood