-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel_evaluate.py
208 lines (165 loc) · 7.32 KB
/
model_evaluate.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
194
195
196
197
198
199
200
201
202
203
204
205
206
207
import numpy as np
import pandas as pd
import sys
import os
import matplotlib.pyplot as plt
import texttable as tt
import tensorflow as tf
from keras.models import Sequential, Model, load_model
from keras.layers import Dense, Conv2D, Flatten, MaxPooling2D, Activation, Input, Dropout, Add
from keras.layers.normalization import BatchNormalization
from keras.callbacks import ModelCheckpoint
from keras.optimizers import Adam
def load_imgs(data_dir):
animal_to_imgs = {}
for animal_name in os.listdir(data_dir):
animal_to_imgs[animal_name] = []
animal_dir = data_dir + "/" + animal_name + "/"
for img_name in os.listdir(animal_dir):
img = plt.imread(animal_dir + img_name)
animal_to_imgs[animal_name].append(img)
return animal_to_imgs
def load_info():
df_classes = pd.read_csv("classes.txt", header=None)
df_predicate_matrix = pd.read_csv("predicate-matrix-binary.txt", header=None)
df_test_classes = pd.read_csv("testclasses.txt", header=None)
df_train_classes = pd.read_csv("trainclasses.txt", header=None)
animal_to_feat = {}
id_to_name, name_to_id = {}, {}
for i, c in enumerate(df_classes[0]):
c_name = c.split()[1]
id_to_name[i] = c_name
name_to_id[c_name] = i
animal_to_feat[c_name] = np.array([int(binary) for binary in df_predicate_matrix.iloc[i, 0].split()])
train_classes, test_classes = [], []
for c in df_train_classes[0]: train_classes.append(c.split()[0])
for c in df_test_classes[0]: test_classes.append(c.split()[0])
return animal_to_feat, id_to_name, name_to_id, train_classes, test_classes
def pred_class(model, img, classes):
s = model.predict(np.expand_dims(img, axis=0))[0]
probs = np.zeros(len(classes))
for i, animal in enumerate(classes):
probs[i] = np.prod(np.abs(s - 1.0 + animal_to_feat[animal]))
return probs.argsort()[-1]
def pred_class_ham(model, img, classes):
s = np.round(model.predict(np.expand_dims(img, axis=0))[0]).astype(int)
score = np.zeros(len(classes))
for i, animal in enumerate(classes):
score[i] = np.sum(np.abs(s - animal_to_feat[animal]))
return score.argsort()[0]
def pred_class_sum(model, img, classes):
s = model.predict(np.expand_dims(img, axis=0))[0]
probs = np.zeros(len(classes))
for i, animal in enumerate(classes):
probs[i] = np.sum(np.abs(s - animal_to_feat[animal]))
return probs.argsort()[0]
def pred_class_harm(model, img, classes):
eps = 1e-5
s = model.predict(np.expand_dims(img, axis=0))[0]
probs = np.zeros(len(classes))
for i, animal in enumerate(classes):
pos = np.sum(np.log(eps + np.abs(s - 1.0 + animal_to_feat[animal])))
neg = np.sum(np.log(eps + np.abs(s - animal_to_feat[animal])))
probs[i] = pos - neg
return probs.argsort()[-1]
def predictions(model, classes, animal_to_images, pred_func=pred_class):
y_pred, y_true = [], []
for i, animal in enumerate(classes):
for img in animal_to_images[animal]:
y_true.append(i)
y_pred.append(pred_func(model, img, classes))
return y_pred, y_true
def pred_features(model, img):
return np.round(model.predict(np.expand_dims(img, axis=0))[0]).astype(int)
def feature_preds(model, classes, animal_to_images):
y_pred, y_true = [], []
for animal in classes:
for img in animal_to_images[animal]:
y_true.append(animal_to_feat[animal])
y_pred.append(pred_features(model, img))
return y_pred, y_true
def draw_table(scores, predicates):
table = tt.Texttable()
table.set_cols_align(["l", "r", "l", "r", "l", "r", "l", "r", "l", "r"])
table.set_cols_valign(["m", "m", "m", "m", "m", "m", "m", "m", "m", "m"])
table.set_cols_width([8, 8, 8, 8, 8, 8, 8, 8, 8, 8])
header = ["Feature", "Score", "Feature", "Score", "Feature", "Score", "Feature", "Score", "Feature", "Score"]
rows = [header]
for i in range(0, len(scores), 5):
temp = []
for j in range(5):
temp.append(predicates[i + j])
temp.append(round(scores[i + j], 3))
rows.append(temp)
table.add_rows(rows)
print(table.draw())
if __name__=="__main__":
print("Load data")
animal_to_imgs = load_imgs("images_128x128")
animal_to_feat, id_to_name, name_to_id, train_classes, test_classes = load_info()
all_classes = train_classes + test_classes
predicate_file = pd.read_csv("predicates.txt", header=None)
predicates = []
for line in predicate_file.iloc[:,0]: predicates.append(line.split()[-1])
print("Load model 1")
model1 = load_model("models/model_1.h5")
print("Load model 2")
model2 = load_model("models/final-model-90.hdf5")
print("Load dropout model")
dropout_model = load_model("models/dropout_model_1.h5")
print("Compute model 1 test accuracy")
y_pred, y_test = predictions(model1, test_classes, animal_to_imgs, pred_func=pred_class_harm)
print(np.mean(np.array(y_pred) == np.array(y_test)))
print("Compute model 2 test accuracy")
y_pred, y_test = predictions(model2, test_classes, animal_to_imgs, pred_func=pred_class_harm)
print(np.mean(np.array(y_pred) == np.array(y_test)))
print("Compute dropout model test accuracy")
y_pred, y_test = predictions(dropout_model, test_classes, animal_to_imgs, pred_func=pred_class_harm)
print(np.mean(np.array(y_pred) == np.array(y_test)))
print("Compute test feature accuracy for 1")
y_pred, y_test = feature_preds(model1, test_classes, animal_to_imgs)
feat_accuracy = np.mean(np.abs(np.array(y_pred) + np.array(y_test) - 1), axis=0)
draw_table(feat_accuracy, predicates)
print("Compute test feature accuracy for 2")
y_pred, y_test = feature_preds(model2, test_classes, animal_to_imgs)
feat_accuracy = np.mean(np.abs(np.array(y_pred) + np.array(y_test) - 1), axis=0)
draw_table(feat_accuracy, predicates)
print("Compute test feature accuracy for dropout model")
y_pred, y_test = feature_preds(dropout_model, test_classes, animal_to_imgs)
print(np.mean(np.abs(np.array(y_pred) + np.array(y_test) - 1)))
print("frequencies!!!")
freqs = np.zeros(len(predicates))
N = 0
for animal, animal_imgs in animal_to_imgs.items():
N += len(animal_imgs)
for i, label in enumerate(animal_to_feat[animal]):
freqs[i] += label * len(animal_imgs)
freqs = freqs / N
draw_table(freqs, predicates)
print("irregularities!")
irregularities = np.zeros(len(test_classes))
for i, animal in enumerate(test_classes):
irregularities[i] = - np.log(np.abs(1.0 - animal_to_feat[animal] - freqs)).sum()
print(animal + ": {}".format(irregularities[i]))
print("Distribution of animals in train classes")
train_dist = np.zeros(len(train_classes))
test_dist = np.zeros(len(test_classes))
for i, animal in enumerate(train_classes): train_dist[i] = len(animal_to_imgs[animal])
for i, animal in enumerate(test_classes): test_dist[i] = len(animal_to_imgs[animal])
train_dist = train_dist / train_dist.sum()
test_dist = test_dist / test_dist.sum()
print("Train distribution:")
for i, animal in enumerate(train_classes):
print(animal + ": {}".format(train_dist[i]))
print("Test distribution:")
for i, animal in enumerate(test_classes):
print(animal + ": {}".format(test_dist[i]))
print("Compute train feature accuracy for 1")
y_pred, y_test = feature_preds(model1, train_classes, animal_to_imgs)
print(np.mean(np.array(y_pred) == np.array(y_test)))
print("Compute train feature accuracy for 2")
y_pred, y_test = feature_preds(model2, train_classes, animal_to_imgs)
print(np.mean(np.array(y_pred) == np.array(y_test)))
print("Compute train feature accuracy for dropout")
y_pred, y_test = feature_preds(dropout_model, train_classes, animal_to_imgs)
print(np.mean(np.array(y_pred) == np.array(y_test)))