astra / metrics.py
suryadev1's picture
fine
5c72fe4
import numpy as np
from scipy.special import softmax
class CELoss(object):
def compute_bin_boundaries(self, probabilities = np.array([])):
#uniform bin spacing
if probabilities.size == 0:
bin_boundaries = np.linspace(0, 1, self.n_bins + 1)
self.bin_lowers = bin_boundaries[:-1]
self.bin_uppers = bin_boundaries[1:]
else:
#size of bins
bin_n = int(self.n_data/self.n_bins)
bin_boundaries = np.array([])
probabilities_sort = np.sort(probabilities)
for i in range(0,self.n_bins):
bin_boundaries = np.append(bin_boundaries,probabilities_sort[i*bin_n])
bin_boundaries = np.append(bin_boundaries,1.0)
self.bin_lowers = bin_boundaries[:-1]
self.bin_uppers = bin_boundaries[1:]
def get_probabilities(self, output, labels, logits):
#If not probabilities apply softmax!
if logits:
self.probabilities = softmax(output, axis=1)
else:
self.probabilities = output
self.labels = labels
self.confidences = np.max(self.probabilities, axis=1)
self.predictions = np.argmax(self.probabilities, axis=1)
self.accuracies = np.equal(self.predictions,labels)
def binary_matrices(self):
idx = np.arange(self.n_data)
#make matrices of zeros
pred_matrix = np.zeros([self.n_data,self.n_class])
label_matrix = np.zeros([self.n_data,self.n_class])
#self.acc_matrix = np.zeros([self.n_data,self.n_class])
pred_matrix[idx,self.predictions] = 1
label_matrix[idx,self.labels] = 1
self.acc_matrix = np.equal(pred_matrix, label_matrix)
def compute_bins(self, index = None):
self.bin_prop = np.zeros(self.n_bins)
self.bin_acc = np.zeros(self.n_bins)
self.bin_conf = np.zeros(self.n_bins)
self.bin_score = np.zeros(self.n_bins)
if index == None:
confidences = self.confidences
accuracies = self.accuracies
else:
confidences = self.probabilities[:,index]
accuracies = self.acc_matrix[:,index]
for i, (bin_lower, bin_upper) in enumerate(zip(self.bin_lowers, self.bin_uppers)):
# Calculated |confidence - accuracy| in each bin
in_bin = np.greater(confidences,bin_lower.item()) * np.less_equal(confidences,bin_upper.item())
self.bin_prop[i] = np.mean(in_bin)
if self.bin_prop[i].item() > 0:
self.bin_acc[i] = np.mean(accuracies[in_bin])
self.bin_conf[i] = np.mean(confidences[in_bin])
self.bin_score[i] = np.abs(self.bin_conf[i] - self.bin_acc[i])
class MaxProbCELoss(CELoss):
def loss(self, output, labels, n_bins = 15, logits = True):
self.n_bins = n_bins
super().compute_bin_boundaries()
super().get_probabilities(output, labels, logits)
super().compute_bins()
#http://people.cs.pitt.edu/~milos/research/AAAI_Calibration.pdf
class ECELoss(MaxProbCELoss):
def loss(self, output, labels, n_bins = 15, logits = True):
super().loss(output, labels, n_bins, logits)
return np.dot(self.bin_prop,self.bin_score)
class MCELoss(MaxProbCELoss):
def loss(self, output, labels, n_bins = 15, logits = True):
super().loss(output, labels, n_bins, logits)
return np.max(self.bin_score)
#https://arxiv.org/abs/1905.11001
#Overconfidence Loss (Good in high risk applications where confident but wrong predictions can be especially harmful)
class OELoss(MaxProbCELoss):
def loss(self, output, labels, n_bins = 15, logits = True):
super().loss(output, labels, n_bins, logits)
return np.dot(self.bin_prop,self.bin_conf * np.maximum(self.bin_conf-self.bin_acc,np.zeros(self.n_bins)))
#https://arxiv.org/abs/1904.01685
class SCELoss(CELoss):
def loss(self, output, labels, n_bins = 15, logits = True):
sce = 0.0
self.n_bins = n_bins
self.n_data = len(output)
self.n_class = len(output[0])
super().compute_bin_boundaries()
super().get_probabilities(output, labels, logits)
super().binary_matrices()
for i in range(self.n_class):
super().compute_bins(i)
sce += np.dot(self.bin_prop,self.bin_score)
return sce/self.n_class
class TACELoss(CELoss):
def loss(self, output, labels, threshold = 0.01, n_bins = 15, logits = True):
tace = 0.0
self.n_bins = n_bins
self.n_data = len(output)
self.n_class = len(output[0])
super().get_probabilities(output, labels, logits)
self.probabilities[self.probabilities < threshold] = 0
super().binary_matrices()
for i in range(self.n_class):
super().compute_bin_boundaries(self.probabilities[:,i])
super().compute_bins(i)
tace += np.dot(self.bin_prop,self.bin_score)
return tace/self.n_class
#create TACELoss with threshold fixed at 0
class ACELoss(TACELoss):
def loss(self, output, labels, n_bins = 15, logits = True):
return super().loss(output, labels, 0.0 , n_bins, logits)