-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathOpenmax.py
193 lines (170 loc) · 8.9 KB
/
Openmax.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
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
from typing import Dict, Any, Union
import numpy as np
from keras import backend as K
import scipy.spatial.distance as spd
import os
from numpy.core._multiarray_umath import ndarray
from tensorflow.keras.models import load_model
from EVT_fitting import weibull_tailfitting
from compute_openmax import recalibrate_scores
from tqdm import tqdm
class Openmax:
"""Class for openmax computations.
Constructor needs a pretrained model with 2 named layers. One should represent the activation vectors and the other
contains the softmax scores. The input data and labels as a tuple, the input shape of the first layer,
optional arguments that represent the observed named layers of the model(eg. dense_01, FNC7, softmax...)
The layer names have to match the layer names of the model
"""
DATA_PATH = "data/openmax/"
MODEL_PATH = "model/openmax/"
def __init__(self, model, x_data, y_data, input_shape: tuple, name: str, overwrite=False, *args):
self.threshold = 0.5
self.weibull_model = None
self.input_shape = (1,) + input_shape
self.model = model
self.x_data = x_data
self.y_data = y_data
self.name = name
self.observations = []
self.backend_functions = []
self.CLASSES = np.unique(self.y_data)
self.overwrite = overwrite
for observation in args:
self.observations.append(observation)
self.init_backend_functions(observation)
if not os.path.exists(self.DATA_PATH+self.name+"/"):
os.makedirs(self.DATA_PATH+self.name+"/")
if not os.path.exists(self.MODEL_PATH):
os.makedirs(self.MODEL_PATH)
def save_model(self):
self.model.save(self.MODEL_PATH + self.name+".h5")
def load_model(self):
self.model = load_model(self.MODEL_PATH + self.name+".h5")
self.backend_functions = []
for observation in self.observations:
self.init_backend_functions(observation)
return self.model
def get_correctly_classified(self):
predictions = self.model.predict(self.x_data)
out = []
for idx, p in enumerate(predictions):
p_label = np.argmax(p)
if p_label == self.y_data[idx] and np.amax(p) > self.threshold:
out.append(True)
else:
out.append(False)
percentage = sum(out)/len(self.x_data)
print("{} percent of the data were predicted correctly and are considered for the probability distribution.".format(percentage*100))
return np.array(out)
def init_backend_functions(self, layer):
if type(layer) is str:
self.backend_functions.append(K.function([self.model.layers[0].input, K.learning_phase()],
[self.model.get_layer(layer).output]))
else:
self.backend_functions.append(K.function([self.model.layers[0].input, K.learning_phase()],
[self.model.layers[layer].output]))
def get_backend_function(self, layer, input):
return self.backend_functions[self.observations.index(layer)]([input, 0])[0]
def compute_feature(self, image_set):
if image_set.shape == self.input_shape:
return np.asarray(self.get_backend_function(self.observations[0], image_set))
out = []
for image in image_set:
out.append(self.get_backend_function(self.observations[0], image.reshape(self.input_shape)))
return np.asarray(out)
def compute_score(self, image_set):
if image_set.shape == self.input_shape:
return np.asarray(self.get_backend_function(self.observations[1], image_set))
out = []
for image in image_set:
out.append(self.get_backend_function(self.observations[1], image.reshape(self.input_shape)))
return np.asarray(out)
def compute_means(self, feature):
return np.mean(feature, axis=0)
def compute_distance(self, mean_train_vector, feature, category_name):
eu_dist, cos_dist, eucos_dist = [], [], []
for feat in feature:
eu_dist += [spd.euclidean(mean_train_vector, feat)]
cos_dist += [spd.cosine(mean_train_vector, feat)]
eucos_dist += [spd.euclidean(mean_train_vector, feat) / 200. +
spd.cosine(mean_train_vector, feat)]
distances = {'eucos': eucos_dist, 'cosine': cos_dist, 'euclidean': eu_dist}
return distances
def prepare_data(self, x, y):
index = np.argsort(y)
return x[index], y[index]
def sort_data(self):
index = self.get_correctly_classified()
x_data = self.x_data[index]
y_data = self.y_data[index]
sorted_x, sorted_y = self.prepare_data(x_data, y_data)
return sorted_x, sorted_y
def compute_data(self):
print("Started Data Computation")
x_data, y_data = self.sort_data()
all_means = []
all_distances = []
for i in tqdm(range(len(self.CLASSES))):
feat = self.compute_feature(x_data[y_data == i])
mean = self.compute_means(feat)
distance = self.compute_distance(mean, feat, i)
all_means.append(mean)
all_distances.append(distance)
np.save(self.DATA_PATH + self.name + "/" + 'activation_vector_class' + str(i), feat)
np.save(self.DATA_PATH + self.name + "/" + 'mean', all_means)
np.save(self.DATA_PATH + self.name + "/" + 'distance', all_distances)
print("Finished Data Computation")
def set_threshold(self, threshold: float):
self.threshold = threshold
def create_weibull(self, distance_type: str, tailsize=20, overwrite=False):
if overwrite:
self.compute_data()
elif len(os.listdir(self.DATA_PATH + self.name+"/")) == 0:
self.compute_data()
means = np.load(self.DATA_PATH + self.name + "/" + 'mean.npy', allow_pickle=True)
distances = np.load(self.DATA_PATH + self.name + "/" + 'distance.npy', allow_pickle=True)
self.weibull_model = weibull_tailfitting(means, distances, range(len(self.CLASSES)),
distance_type=distance_type, tailsize=tailsize)
return self.weibull_model
def predict(self, image, distance_type='eucos', alpha=10):
if self.weibull_model is None:
self.create_weibull(distance_type, self.overwrite)
if not image.shape == self.input_shape:
image = image.reshape(self.input_shape)
imgarr = {'scores': self.compute_score(image),
self.observations[0]: self.compute_feature(image)}
openmax, softmax = recalibrate_scores(self.weibull_model, range(len(self.CLASSES)), imgarr,
self.observations[0], distance_type=distance_type,
classes=len(self.CLASSES),alpharank=alpha)
return openmax, softmax
def adapt_params(self, distance_type='eucos', tail_range=range(0,40), alpha_range=range(0,10)):
adaption = np.zeros((len(alpha_range) + 1, len(tail_range) + 1))
for image in tqdm(self.x_data):
for alpha in alpha_range:
for tail in tail_range:
self.create_weibull(distance_type, tailsize=tail, overwrite=False)
if not image.shape == self.input_shape:
image = image.reshape(self.input_shape)
imgarr = {'scores': self.compute_score(image),
self.observations[0]: self.compute_feature(image)}
openmax, softmax = recalibrate_scores(self.weibull_model, range(len(self.CLASSES)), imgarr,
self.observations[0], distance_type=distance_type,
classes=len(self.CLASSES), alpharank=alpha)
if np.argmax(openmax) == np.argmax(softmax):
adaption[alpha][tail] += 1
return adaption
def adapt_tailsize(self, distance_type='eucos', adaption_range=range(1, 40)):
tails = np.zeros(len(adaption_range)+1)
for image in self.x_data:
for i in adaption_range:
self.create_weibull(distance_type, tailsize=i, overwrite=False)
if not image.shape == self.input_shape:
image = image.reshape(self.input_shape)
imgarr = {'scores': self.compute_score(image),
self.observations[0]: self.compute_feature(image)}
openmax, softmax = recalibrate_scores(self.weibull_model, range(len(self.CLASSES)), imgarr,
self.observations[0], distance_type=distance_type,
classes=len(self.CLASSES))
if np.argmax(openmax) == np.argmax(softmax):
tails[i] += 1
return tails