-
Notifications
You must be signed in to change notification settings - Fork 1
/
experiments_SVGP.py
308 lines (273 loc) · 18.4 KB
/
experiments_SVGP.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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
"""
Copyright (C) 2022 Cognizant Digital Business, Evolutionary AI. All Rights Reserved.
Issued under the Academic Public License.
You can be released from the terms, and requirements of the Academic public license by purchasing a commercial license.
"""
from __future__ import absolute_import, division, print_function
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import pickle
import os
import time
from util import load_UCI121, dataset_read, RIO_variants_running, RIO_MRBF_running, RIO_MRBF_multiple_running
import numpy as np
from sklearn.metrics import mean_absolute_error
from scipy.special import softmax
import scipy
import trustscore
#main file to run tests for all the RIO variants on all the datasets
print(tf.__version__)
model_name = "SVGP"
#number of Epochs for NN training
EPOCHS = 1000
#number of inducing points for SVGP
M = 50
dataset_name_list = ["balance-scale", "blood", "abalone", "annealing", "car", "contrac", "mammographic", "miniboone",
"wine", "lenses","breast-cancer-wisc-prog","haberman-survival","post-operative","spectf","plant-texture",
"pima","synthetic-control","iris","breast-tissue","conn-bench-vowel-deterding","ozone","oocytes_trisopterus_states_5b",
"twonorm","audiology-std","heart-switzerland","musk-2","spambase","lung-cancer","molec-biol-promoter","congressional-voting",
"conn-bench-sonar-mines-rocks","breast-cancer-wisc-diag","thyroid","spect","optical","arrhythmia","oocytes_merluccius_nucleus_4d",
"credit-approval", "cylinder-bands", "energy-y1", "energy-y2", "hill-valley", "image-segmentation", "led-display", "magic",
"cardiotocography-3clases", "chess-krvk", "chess-krvkp", "connect-4",
"Phishing","messidor","Bioconcentration","Climate","yeast",
"adult", "bank", "cardiotocography-10clases",
"nursery","oocytes_trisopterus_nucleus_2f","low-res-spect","ilpd-indian-liver","statlog-image","flags","semeion",
"wall-following","soybean","zoo","hayes-roth","plant-margin","hepatitis","wine-quality-red","parkinsons","wine-quality-white","mushroom",
"monks-3","breast-cancer","pittsburg-bridges-REL-L","statlog-heart","statlog-landsat","fertility","monks-1","statlog-vehicle",
"vertebral-column-3clases","ionosphere","pittsburg-bridges-TYPE","acute-nephritis","libras","horse-colic","oocytes_merluccius_states_2f","breast-cancer-wisc",
"pittsburg-bridges-MATERIAL","statlog-shuttle","waveform","steel-plates","statlog-german-credit","trains","statlog-australian-credit",
"acute-inflammation","page-blocks","molec-biol-splice","seeds","titanic","ringnorm","musk-1","glass","pittsburg-bridges-T-OR-D",
"planning","dermatology","monks-2","ecoli","primary-tumor","waveform-noise","teaching","lymphography","balloons","heart-cleveland",
"pendigits","plant-shape","letter","tic-tac-toe","echocardiogram","vertebral-column-2clases","heart-va","heart-hungarian","pittsburg-bridges-SPAN"]
# For newly added datasets only
new_dataset_name_list = ["Phishing","messidor","Bioconcentration","Climate"]
new_label_name_list = ["Result", "Class", "Class", "outcome"]
new_minibatch_size_list = [1082,921,623,432]
new_num_class_list = [3,2,3,2]
new_dataset_index_dict = {}
for i in range(len(new_dataset_name_list)):
new_dataset_index_dict[new_dataset_name_list[i]] = i
def build_regression_model(layer_width, input_dim):
model = keras.Sequential([
layers.Dense(layer_width, activation=tf.nn.relu, input_shape=[input_dim]),
layers.Dense(layer_width, activation=tf.nn.relu),
layers.Dense(1)
])
optimizer = tf.keras.optimizers.RMSprop(0.001)
model.compile(loss='mean_squared_error',
optimizer=optimizer,
metrics=['mean_absolute_error', 'mean_squared_error'])
return model
def build_classification_model(layer_width, num_class, input_dim):
model = keras.Sequential([
layers.Dense(layer_width, activation=tf.nn.relu, input_shape=[input_dim]),
layers.Dense(layer_width, activation=tf.nn.relu),
layers.Dense(num_class)
])
optimizer = tf.keras.optimizers.RMSprop(0.001)
model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer="adam",#optimizer,
metrics=['accuracy'])
return model
def one_hot_encoding(origin_labels, num_class):
one_hot_labels = np.zeros((len(origin_labels),num_class))
one_hot_labels[np.arange(len(origin_labels)),origin_labels] = 1
return one_hot_labels
def acc_calculate(predictions, labels):
prediction_class = np.argmax(predictions, axis=1)
num_correct = np.sum(prediction_class==labels)
acc = num_correct/len(labels)
return acc
def run_RIO_classification(framework_variant, kernel_type, M, rio_data, rio_setups, algo_spec):
mean_list = []
var_list = []
correction_list = []
NN_MAE_list = []
RIO_MAE_list = []
PCT_within95Interval_list = []
PCT_within90Interval_list = []
PCT_within68Interval_list = []
computation_time_list = []
hyperparameter_list = []
num_optimizer_iter_list = []
if algo_spec == "moderator_direct_target":
train_labels_class = rio_data["one_hot_train_labels"][:,0].copy()
test_labels_class = rio_data["one_hot_test_labels"][:,0].copy()
train_NN_predictions_class = rio_data["one_hot_train_labels"][:,0].copy()
test_NN_predictions_class = rio_data["one_hot_test_labels"][:,0].copy()
for i in range(len(train_labels_class)):
train_labels_class[i] = np.max(rio_data["train_NN_predictions_softmax"][i])
train_NN_predictions_class[i] = np.max(rio_data["train_NN_predictions_softmax"][i])
if rio_data["train_check"][i]:
train_labels_class[i] = 1.0
else:
train_labels_class[i] = 0.0
for i in range(len(test_labels_class)):
test_labels_class[i] = np.max(rio_data["test_NN_predictions_softmax"][i])
test_NN_predictions_class[i] = np.max(rio_data["test_NN_predictions_softmax"][i])
if rio_data["test_check"][i]:
test_labels_class[i] = 1.0
else:
test_labels_class[i] = 0.0
train_NN_predictions_all = rio_data["train_NN_predictions_softmax"]
test_NN_predictions_all = rio_data["test_NN_predictions_softmax"]
NN_MAE = mean_absolute_error(test_labels_class, test_NN_predictions_class)
if framework_variant == "GP_corrected" or framework_variant == "GP":
with tf.Graph().as_default() as tf_graph, tf.Session(graph=tf_graph).as_default():
MAE, PCT_within95Interval, PCT_within90Interval, PCT_within68Interval, mean, var, computation_time, hyperparameter, num_optimizer_iter, mean_train, var_train = RIO_MRBF_multiple_running(framework_variant, \
kernel_type, \
rio_data["normed_train_data"], \
rio_data["normed_test_data"], \
train_labels_class, \
test_labels_class, \
train_NN_predictions_class, \
test_NN_predictions_class, \
train_NN_predictions_all, \
test_NN_predictions_all, \
M, \
rio_setups["use_ard"], \
rio_setups["scale_array"], \
rio_setups["separate_opt"])
else:
with tf.Graph().as_default() as tf_graph, tf.Session(graph=tf_graph).as_default():
MAE, PCT_within95Interval, PCT_within90Interval, PCT_within68Interval, mean, var, computation_time, hyperparameter, num_optimizer_iter, mean_train, var_train = RIO_variants_running(framework_variant, \
kernel_type, \
rio_data["normed_train_data"], \
rio_data["normed_test_data"], \
train_labels_class, \
test_labels_class, \
train_NN_predictions_class, \
test_NN_predictions_class, \
M, \
rio_setups["use_ard"], \
rio_setups["scale_array"])
if framework_variant == "GP_corrected" or framework_variant == "GP_corrected_inputOnly" or framework_variant == "GP_corrected_outputOnly" or algo_spec == "moderator_residual_target":
correction_list.append(mean)
mean_list.append(mean+test_NN_predictions_class)
correction = mean.copy()
mean = mean+test_NN_predictions_class
else:
mean_list.append(mean)
correction_list.append(mean)
correction = mean.copy()
var_list.append(var)
NN_MAE_list.append(NN_MAE)
RIO_MAE_list.append(MAE)
PCT_within95Interval_list.append(PCT_within95Interval)
PCT_within90Interval_list.append(PCT_within90Interval)
PCT_within68Interval_list.append(PCT_within68Interval)
computation_time_list.append(computation_time)
hyperparameter_list.append(hyperparameter)
num_optimizer_iter_list.append(num_optimizer_iter)
correction_list_transpose = np.array(correction_list).transpose()
mean_list_transpose = np.array(mean_list).transpose()
var_list_transpose = np.array(var_list).transpose()
print("mean of True: {}".format(np.mean(mean[np.where(rio_data["test_check"])])))
print("mean of False: {}".format(np.mean(mean[np.where(rio_data["test_check"] == False)])))
exp_result = {}
exp_result["mean"] = mean
exp_result["var"] = var
exp_result["RIO_MAE"] = MAE
exp_result["PCT_within95Interval"] = PCT_within95Interval
exp_result["PCT_within90Interval"] = PCT_within90Interval
exp_result["PCT_within68Interval"] = PCT_within68Interval
exp_result["computation_time"] = computation_time
exp_result["hyperparameter"] = hyperparameter
exp_result["num_optimizer_iter"] = num_optimizer_iter
exp_result["test_labels"] = rio_data["test_labels"].values.reshape(-1)
exp_result["test_NN_predictions"] = rio_data["test_NN_predictions"]
exp_result["mean_train"] = mean_train
exp_result["var_train"] = var_train
exp_result["train_labels"] = rio_data["train_labels"].values.reshape(-1)
exp_result["train_NN_predictions"] = rio_data["train_NN_predictions"]
exp_result["mean_correct_train"] = np.mean(mean_train[np.where(rio_data["train_check"])])
exp_result["mean_incorrect_train"] = np.mean(mean_train[np.where(rio_data["train_check"] == False)])
exp_result["mean_correct_test"] = np.mean(mean[np.where(rio_data["test_check"])])
exp_result["mean_incorrect_test"] = np.mean(mean[np.where(rio_data["test_check"] == False)])
return exp_result
for dataset_index in range(len(dataset_name_list)):
dataset_name = dataset_name_list[dataset_index]
NN_size = "64+64"
layer_width = 64
RUNS = 10
NN_info = NN_size
if dataset_name in new_dataset_name_list:
label_name = new_label_name_list[new_dataset_index_dict[dataset_name]]
minibatch_size = new_minibatch_size_list[new_dataset_index_dict[dataset_name]]
num_class = new_num_class_list[new_dataset_index_dict[dataset_name]]
dataset = dataset_read(dataset_name)
else:
normed_dataset, labels = load_UCI121(dataset_name)
num_class = np.max(labels.values)+1
print("num_class: {}".format(num_class))
for run in range(RUNS):
print("run{} start".format(run))
tf.reset_default_graph()
with tf.Session(graph=tf.Graph()):
# preprocess data
if dataset_name in new_dataset_name_list:
train_dataset = dataset.sample(frac=0.8,random_state=run)
test_dataset = dataset.drop(train_dataset.index)
train_labels = train_dataset.pop(label_name).astype(int)
test_labels = test_dataset.pop(label_name).astype(int)
train_stats = train_dataset.describe()
train_stats = train_stats.transpose()
normed_train_data = (train_dataset - train_stats['mean']) / train_stats['std']
normed_test_data = (test_dataset - train_stats['mean']) / train_stats['std']
else:
normed_train_data = normed_dataset.sample(frac=0.8,random_state=run)
normed_test_data = normed_dataset.drop(normed_train_data.index)
train_labels = labels.take(normed_train_data.index)
test_labels = labels.drop(normed_train_data.index)
minibatch_size = len(normed_train_data)
# load data for one trial
algo_spec = "moderator_residual_target"
kernel_type = "RBF+RBF"
framework_variant = "GP"
add_info = "+separate_opt"
trial = 0
dir_name = "Results"
result_file_name = os.path.join(os.path.dirname(os.path.abspath(__file__)),dir_name,'{}_exp_result_{}_{}_{}_run{}_trail{}.pkl'.format(dataset_name, framework_variant, kernel_type, algo_spec+add_info, run, trial))
with open(result_file_name, 'rb') as result_file:
exp_result = pickle.load(result_file)
test_NN_predictions = exp_result["test_NN_predictions"]
train_NN_predictions = exp_result["train_NN_predictions"]
one_hot_train_labels = one_hot_encoding(train_labels.values.reshape(-1), num_class)
one_hot_test_labels = one_hot_encoding(test_labels.values.reshape(-1), num_class)
# This feature not used in RED
scale_correct = 1.0 #len(train_labels.values)/(2*num_correct)
scale_incorrect = 1.0 #len(train_labels.values)/(2*num_incorrect)
scale_array = np.ones(len(train_labels.values))
if scale_correct < scale_incorrect:
for k in range(len(train_labels.values)):
if train_NN_correct[k]:
scale_array[k] = scale_correct
else:
scale_array[k] = scale_incorrect
rio_data = {}
rio_setups = {}
rio_data["normed_train_data"] = normed_train_data
rio_data["normed_test_data"] = normed_test_data
rio_data["train_NN_predictions"] = train_NN_predictions
rio_data["test_NN_predictions"] = test_NN_predictions
rio_data["train_labels"] = train_labels
rio_data["test_labels"] = test_labels
rio_data["train_NN_predictions_softmax"] = softmax(exp_result["train_NN_predictions"], axis=1)
rio_data["test_NN_predictions_softmax"] = softmax(exp_result["test_NN_predictions"], axis=1)
rio_data["train_check"] = (rio_data["train_labels"].values.reshape(-1)==np.argmax(rio_data["train_NN_predictions"], axis=1))
rio_data["test_check"] = (rio_data["test_labels"].values.reshape(-1)==np.argmax(rio_data["test_NN_predictions"], axis=1))
rio_data["one_hot_train_labels"] = one_hot_train_labels
rio_data["one_hot_test_labels"] = one_hot_test_labels
rio_setups["use_ard"] = True
rio_setups["scale_array"] = scale_array
rio_setups["separate_opt"] = False
kernel_type = "RBF"
framework_variant = "GP_inputOnly"
algo_spec = "moderator_direct_target"
add_info = ""
exp_result = run_RIO_classification(framework_variant, kernel_type, M, rio_data, rio_setups, algo_spec)
result_file_name = os.path.join(os.path.dirname(os.path.abspath(__file__)),'Results','{}_exp_result_{}_{}_{}_run{}.pkl'.format(dataset_name, framework_variant, kernel_type, algo_spec+add_info, run))
with open(result_file_name, 'wb') as result_file:
pickle.dump(exp_result, result_file)