/
calibrate.py
130 lines (101 loc) · 4.18 KB
/
calibrate.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
from sacred import Experiment
EXPERIMENT_NAME = "calibrate"
ex = Experiment(EXPERIMENT_NAME, ingredients=[])
# https://github.com/gpleiss/temperature_scaling/blob/master/temperature_scaling.py
class ECELoss(nn.Module):
"""
Calculates the Expected Calibration Error of a model.
(This isn't necessary for temperature scaling, just a cool metric).
The input to this loss is the logits of a model, NOT the softmax scores.
This divides the confidence outputs into equally-sized interval bins.
In each bin, we compute the confidence gap:
bin_gap = | avg_confidence_in_bin - accuracy_in_bin |
We then return a weighted average of the gaps, based on the number
of samples in each bin
See: Naeini, Mahdi Pakdaman, Gregory F. Cooper, and Milos Hauskrecht.
"Obtaining Well Calibrated Probabilities Using Bayesian Binning." AAAI.
2015.
"""
def __init__(self, n_bins=15):
"""
n_bins (int): number of confidence interval bins
"""
super(ECELoss, self).__init__()
bin_boundaries = torch.linspace(0, 1, n_bins + 1)
self.bin_lowers = bin_boundaries[:-1]
self.bin_uppers = bin_boundaries[1:]
def forward(self, logits, labels):
softmaxes = F.softmax(logits, dim=1)
confidences, predictions = torch.max(softmaxes, 1)
accuracies = predictions.eq(labels)
ece = torch.zeros(1, device=logits.device)
for bin_lower, bin_upper in zip(self.bin_lowers, self.bin_uppers):
# Calculated |confidence - accuracy| in each bin
in_bin = confidences.gt(bin_lower.item()) * confidences.le(bin_upper.item())
prop_in_bin = in_bin.float().mean()
if prop_in_bin.item() > 0:
accuracy_in_bin = accuracies[in_bin].float().mean()
avg_confidence_in_bin = confidences[in_bin].mean()
ece += torch.abs(avg_confidence_in_bin - accuracy_in_bin) * prop_in_bin
return ece
def plot_calibration_error(logits, targets):
confidences = F.softmax(logits, -1).max(-1).values.detach().numpy()
accuracies = logits.argmax(-1).eq(targets).numpy()
print(confidences)
print(accuracies)
n_bins = 15
bin_boundaries = np.linspace(0, 1, n_bins + 1)
bin_lowers = bin_boundaries[:-1]
bin_uppers = bin_boundaries[1:]
ece = 0.0
max_err = 0.0
plot_acc = []
for bin_lower, bin_upper in zip(bin_lowers, bin_uppers):
in_bin = (confidences > bin_lower) * (confidences <= bin_upper)
prop_in_bin = in_bin.astype(np.float32).mean()
if prop_in_bin > 0.0:
accuracy_in_bin = accuracies[in_bin].astype(np.float32).mean()
avg_confidence_in_bin = confidences[in_bin].mean()
ece += np.abs(avg_confidence_in_bin - accuracy_in_bin) * prop_in_bin
if np.abs(avg_confidence_in_bin - accuracy_in_bin) > max_err:
max_err = np.abs(avg_confidence_in_bin - accuracy_in_bin)
plot_acc.append(accuracy_in_bin)
else:
plot_acc.append(0.0)
plt.bar(
bin_lowers, plot_acc, bin_uppers[0], align="edge", linewidth=1, edgecolor="k"
)
plt.plot([0.0, 1.0], [0.0, 1.0], c="orange", lw=2)
plt.text(
0.02,
0.93,
"acc: {:0.4f}\nece: {:0.4f}\nmce: {:0.4f}".format(
accuracies.astype(np.float32).mean(), ece, max_err
),
fontsize=16,
)
plt.xlabel("confidence")
plt.ylabel("accuracy")
plt.savefig("temp.png", bbox_inches="tight")
@ex.config
def get_config():
# where evaluation runs are located
run_dir = "runs/test"
job_id = None
@ex.automain
def main(run_dir, job_id, _run):
assert job_id is not None, "must specify a job id"
results = torch.load(
os.path.join(run_dir, str(job_id), "results.pth"), map_location="cpu"
)
logits = torch.cat([result["stats"]["logits"] for result in results], 0)
targets = torch.cat([result["stats"]["targets"] for result in results], 0)
ece_module = ECELoss()
print(ece_module.forward(logits, targets))
plot_calibration_error(logits, targets)