-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathunivariate_statistical_testing_comparison.py
177 lines (140 loc) · 7.69 KB
/
univariate_statistical_testing_comparison.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
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import numpy as np
import os
import seaborn as sns
import shutil
# ========== HYPERPARAMETERS ==========
# ***RESEARCHER BEWARE***
# Total number of GANs =
# |NUM_SAMPLES_AVAILABLE_TO_MODEL| * NUM_MODELS_TO_TRAIN_PER_SAMPLE_SIZE * |UNIVARIATE_DISTRIBUTIONS| * 2
NUM_SAMPLES_AVAILABLE_TO_MODEL = np.geomspace(10, 250, num=20)
NUM_MODELS_TO_TRAIN_PER_SAMPLE_SIZE = 5
NUM_SYN_SAMPLES_TO_GENERATE = 25000
UNIVARIATE_DISTRIBUTIONS = ['gaussian_0', 'gaussian_0_1', 'chi_square_9', 'exp_9', 'gaussian_mixture'] # 'gaussian_1'
# ========== OUTPUT DIRECTORIES ==========
OUTPUT_DIR = 'OUTPUT/'
MODELS_OUTPUT_DIR = OUTPUT_DIR + 'MODELS/'
SYN_DATA_OUTPUT_DIR = OUTPUT_DIR + 'SYN_DATA/'
REAL_DATA_OUTPUT_DIR = OUTPUT_DIR + 'REAL_DATA/'
POWER_DIR = OUTPUT_DIR + 'POWER/'
RESULTS_DIR = 'RESULTS/'
shutil.rmtree(OUTPUT_DIR, ignore_errors=True)
os.makedirs(MODELS_OUTPUT_DIR)
os.makedirs(SYN_DATA_OUTPUT_DIR)
os.makedirs(REAL_DATA_OUTPUT_DIR)
os.makedirs(POWER_DIR)
os.makedirs(RESULTS_DIR)
# ========== RUN PIPELINE ==========
def output_dirs(dist, n, k):
model_tag_base = '[{0}]_[n={1}]_[k={2}]'.format(dist, n, k)
model_1_tag = model_tag_base + '_[v=1]'
model_2_tag = model_tag_base + '_[v=2]'
model_1_dir = '{0}{1}/'.format(MODELS_OUTPUT_DIR, model_1_tag)
model_2_dir = '{0}{1}/'.format(MODELS_OUTPUT_DIR, model_2_tag)
syn_data_1_dir = '{0}{1}/'.format(SYN_DATA_OUTPUT_DIR, model_1_tag)
syn_data_2_dir = '{0}{1}/'.format(SYN_DATA_OUTPUT_DIR, model_2_tag)
real_data_1_dir = '{0}{1}/'.format(REAL_DATA_OUTPUT_DIR, model_1_tag)
real_data_2_dir = '{0}{1}/'.format(REAL_DATA_OUTPUT_DIR, model_2_tag)
return model_1_dir, model_2_dir, syn_data_1_dir, syn_data_2_dir, real_data_1_dir, real_data_2_dir
def train_and_generate_samples(num_samples_available_to_model):
# Generate real and synthetic samples:
for k in range(NUM_MODELS_TO_TRAIN_PER_SAMPLE_SIZE):
for dist in UNIVARIATE_DISTRIBUTIONS:
# Set up output directories
model_1_dir, model_2_dir, syn_data_1_dir, syn_data_2_dir, real_data_1_dir, real_data_2_dir = output_dirs(dist, num_samples_available_to_model, k)
# Set up commands
generate_real_cmd_1 = 'python3 sample_prob_dist.py {0} {1} {2}'.format(dist, num_samples_available_to_model, real_data_1_dir)
generate_real_cmd_2 = 'python3 sample_prob_dist.py {0} {1} {2}'.format(dist, num_samples_available_to_model, real_data_2_dir)
train_cmd_1 = 'python3 train_prob_gan.py {0} {1}'.format(real_data_1_dir+'data.npy', model_1_dir)
train_cmd_2 = 'python3 train_prob_gan.py {0} {1}'.format(real_data_2_dir+'data.npy', model_2_dir)
generate_syn_cmd_1 = 'python3 generate_prob_gan.py {0} {1} {2}'.format(model_1_dir + 'generator', NUM_SYN_SAMPLES_TO_GENERATE, syn_data_1_dir)
generate_syn_cmd_2 = 'python3 generate_prob_gan.py {0} {1} {2}'.format(model_2_dir + 'generator', NUM_SYN_SAMPLES_TO_GENERATE, syn_data_2_dir)
# Run commands
os.system(generate_real_cmd_1)
os.system(generate_real_cmd_2)
os.system(train_cmd_1)
os.system(train_cmd_2)
os.system(generate_syn_cmd_1)
os.system(generate_syn_cmd_2)
def compute_power_between_distributions(num_samples_available_to_model, dist_1, dist_2):
t_real_power = []
t_syn_power = []
for k in range(NUM_MODELS_TO_TRAIN_PER_SAMPLE_SIZE):
# Retrieve data directories
# Note syn_data_1_dir contains data generated by v=1 model.
# syn_data_2_dir contains data generated by v=2 model.
# This strategy ensures that null tests aren't testing two synthetic samples
# that were generated by the same GAN.
_, _, syn_data_1_dir, _, real_data_1_dir, _ = output_dirs(dist_1, num_samples_available_to_model, k)
_, _, _, syn_data_2_dir, _, real_data_2_dir = output_dirs(dist_2, num_samples_available_to_model, k)
# Set up compute_power_cmd
real_dataset_1 = real_data_1_dir + 'data.npy'
syn_dataset_1 = syn_data_1_dir + 'data.npy'
real_dataset_2 = real_data_2_dir + 'data.npy'
syn_dataset_2 = syn_data_2_dir + 'data.npy'
power_dir = '{0}[{1}*{2}]_[n={3}]_[k={4}]/'.format(POWER_DIR, dist_1, dist_2, num_samples_available_to_model, k)
compute_power_cmd = 'python3 compute_univariate_power.py {0} {1} {2} {3} {4}'.format(real_dataset_1, syn_dataset_1, real_dataset_2, syn_dataset_2, power_dir)
# Run power computation
os.system(compute_power_cmd)
# Collect results
results = open(power_dir + 'results.txt').readlines()[0].split(',')
t_real_power.append(float(results[0]))
t_syn_power.append(float(results[1]))
# Save power results:
results_pth = '{0}[{1}*{2}]/'.format(RESULTS_DIR, dist_1, dist_2)
real_results_pth = results_pth + 'real.npy'
syn_results_pth = results_pth + 'syn.npy'
if not os.path.exists(results_pth):
os.makedirs(results_pth)
t_real_power_for_sample_size_for_dist1_dist2 = []
t_syn_power_for_sample_size_for_dist1_dist2 = []
else:
t_real_power_for_sample_size_for_dist1_dist2 = np.load(real_results_pth).tolist()
t_syn_power_for_sample_size_for_dist1_dist2 = np.load(syn_results_pth).tolist()
t_real_power_for_sample_size_for_dist1_dist2.append(t_real_power)
t_syn_power_for_sample_size_for_dist1_dist2.append(t_syn_power)
np.save(real_results_pth, np.array(t_real_power_for_sample_size_for_dist1_dist2))
np.save(syn_results_pth, np.array(t_syn_power_for_sample_size_for_dist1_dist2))
def compute_all_power_tests(num_samples_available_to_model):
# For every combination of distributions, compute the power
# of a t test distinguishing between real and synthetic samples respectively
# and save the results.
for i in range(len(UNIVARIATE_DISTRIBUTIONS)):
for j in range(i, len(UNIVARIATE_DISTRIBUTIONS)):
dist_1 = UNIVARIATE_DISTRIBUTIONS[i]
dist_2 = UNIVARIATE_DISTRIBUTIONS[j]
compute_power_between_distributions(num_samples_available_to_model, dist_1, dist_2)
def clear_output_dirs():
shutil.rmtree(MODELS_OUTPUT_DIR)
shutil.rmtree(SYN_DATA_OUTPUT_DIR)
shutil.rmtree(REAL_DATA_OUTPUT_DIR)
shutil.rmtree(POWER_DIR)
# ========== MAIN ==========
for i in range(NUM_SAMPLES_AVAILABLE_TO_MODEL.shape[0]):
n = int(NUM_SAMPLES_AVAILABLE_TO_MODEL[i])
train_and_generate_samples(n)
compute_all_power_tests(n)
clear_output_dirs()
# ========== VISUALIZATION ==========
for i in range(len(UNIVARIATE_DISTRIBUTIONS)):
for j in range(i, len(UNIVARIATE_DISTRIBUTIONS)):
dist_1 = UNIVARIATE_DISTRIBUTIONS[i]
dist_2 = UNIVARIATE_DISTRIBUTIONS[j]
results_pth = '{0}[{1}*{2}]/'.format(RESULTS_DIR, dist_1, dist_2)
real_results_pth = results_pth + 'real.npy'
syn_results_pth = results_pth + 'syn.npy'
t_real_power_for_sample_size_for_dist1_dist2 = np.load(real_results_pth).T
t_syn_power_for_sample_size_for_dist1_dist2 = np.load(syn_results_pth).T
plt.figure()
sns.tsplot(data=t_real_power_for_sample_size_for_dist1_dist2, time=NUM_SAMPLES_AVAILABLE_TO_MODEL, ci=[68, 95], color='blue', condition='T Test Real')
sns.tsplot(data=t_syn_power_for_sample_size_for_dist1_dist2, time=NUM_SAMPLES_AVAILABLE_TO_MODEL, ci=[68, 95], color='orange', condition='T Test Syn')
plt.title('{0} vs {1}'.format(dist_1, dist_2))
plt.xlabel('Real Samples')
plt.ylabel('Power')
plt.ylim([-0.1, 1.1])
plt.legend(loc="upper right")
plt.tight_layout()
plt.savefig('{0}true_sample_size_vs_power.png'.format(results_pth))
plt.close()