/
Addition_EXP.py
256 lines (215 loc) · 10.7 KB
/
Addition_EXP.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
import numpy as np
import learning
import synthetic_data
import pickle
import os
import sys
from shutil import copyfile
import time
import matplotlib.pyplot as plt
import argparse
def tic():
return time.clock()
def toc(t):
return time.clock() - t
def generate_data_simple_addition(N_samples, seq_length, noise=0.):
X = []
Y = []
for i in range(N_samples):
if seq_length > 0:
X.append(np.hstack((np.random.normal(0, 1, [seq_length, 2]),np.ones([seq_length,1]))))
y = X[-1][:,0].sum()-X[-1][:,1].sum() + np.random.normal(0,noise)
else:
X.append([])
y = 0
Y.append(y)
return np.asarray(X),np.asarray(Y).squeeze()
def sum_untill_length(length, dim):
temp = 0
for i in range(1, length+1):
temp+= dim**i
return temp
if __name__ == '__main__':
'''
python Addition_EXP.py 'launch' './new_examples_add' 0.1 2 0.1
'''
L_num_examples = [20, 40, 80, 160, 320, 640, 1500, 2560, 5000]
N_runs = 1
#print(N_runs)
length = 2
test_length = 6
methods = ['NuclearNorm', 'TIHT', 'IHT', 'OLS', 'LSTM']
TIHT_epsilon = 1e-15
TIHT_learning_rate = 1e-1
TIHT_max_iters = 5000
xp_path = './SP_Addition/'
b2 = 100
lr2 = 0.001
epo2 = 1000
tol = 50
verbose = False
parser = argparse.ArgumentParser()
parser.add_argument('-lne', '--list_number_examples', nargs = '+', help='list of examples numbers', type=int)
parser.add_argument('-nr', '--number_runs', help='number of runs', type=int)
parser.add_argument('-le', '--length', help='minimum training length', type=int)
parser.add_argument('-tle', '--testing_length', help='testing length', type=int)
parser.add_argument('-lm', '--method_list', nargs='+', help="List of methods to use")
parser.add_argument('-eps', '--HT_epsilon', help='epsilon for TIHT and IHT', type=float)
parser.add_argument('-lr', '--HT_learning_rate', help='learning rate for TIHT and IHT', type=float)
parser.add_argument('-mi', '--HT_max_iter', help='number of max iterations for TIHT and IHT', type=int)
parser.add_argument('-xp', '--xp_path', help='experiment folder path')
parser.add_argument('-var', '--noise', help='variance of the gaussian noise', type=float)
parser.add_argument('-ns', '--states_number', help='number of states for the model', type=int)
parser.add_argument('-a', '--alpha', help='hyperparameter for nuclear norm method', type=float)
parser.add_argument('-ld', '--load_data', help='load the previously created data', action='store_true')
parser.add_argument('-lr2', help='learning rate for sgd 2rnn', type=float)
parser.add_argument('-epo2', help='number of epochs for sgd 2rnn', type=int)
parser.add_argument('-b2', '--batch_size', help='batch size for sgd 2rnn', type=int)
parser.add_argument('-t', '--tolerance', help='tolerance for sgd 2rnn', type=int)
args = parser.parse_args()
if args.noise != None:
noise_level = args.noise
else:
raise Exception('Did not initialize noise_level, try set up after -var argument')
if args.states_number != None:
num_states = args.states_number
else:
raise Exception('Did not initialize state numbers, try set up after -ns argument')
if args.alpha != None:
alpha = args.alpha
else:
raise Exception('Did not initialize alpha, try set up after -a argument')
load_data = True
if args.list_number_examples:
L_num_examples = args.list_number_examples
if args.number_runs:
N_runs = args.number_runs
if args.length:
length = args.length
if args.testing_length:
test_length = args.testing_length
if args.method_list:
methods = args.method_list
if args.HT_epsilon:
TIHT_epsilon = args.HT_epsilon
if args.HT_learning_rate:
TIHT_learning_rate = args.HT_learning_rate
if args.HT_max_iter:
TIHT_max_iters = args.HT_max_iter
if args.xp_path:
xp_path = args.xp_path
if args.lr2:
lr2 = args.lr2
if args.epo2:
epo2 = args.epo2
if args.batch_size:
b2 = args.batch_size
if args.tolerance:
tol = args.tolerance
if not os.path.exists(xp_path):
os.makedirs(xp_path)
if not os.path.exists(xp_path + 'noise_' + str(noise_level)):
os.makedirs(xp_path + 'noise_' + str(noise_level))
xp_path = xp_path + 'noise_' + str(noise_level)+'/'
if not os.path.exists(xp_path):
os.makedirs(xp_path)
results = dict([(m, {}) for m in methods])
for num_examples in L_num_examples:
for m in methods:
results[m][num_examples] = []
results['NUM_EXAMPLES'] = L_num_examples
times = dict([(m, {}) for m in methods])
for num_examples in L_num_examples:
for m in methods:
times[m][num_examples] = []
times['NUM_EXAMPLES'] = L_num_examples
for run in range(N_runs):
if load_data == False:
data_function = lambda l: generate_data_simple_addition(1000, l, noise=noise_level)
Xtest, ytest = data_function(test_length)
with open('./Data/Addition/noise_' + str(noise_level) + '/Test.pickle', 'wb') as f:
pickle.dump([Xtest, ytest], f)
elif load_data == True:
data_function = lambda l: generate_data_simple_addition(1000, l, noise=noise_level)
with open('./Data/Addition/noise_' + str(noise_level) + '/Test.pickle', 'rb') as f:
[Xtest, ytest] = pickle.load(f)
print("test MSE of zero function", np.mean(ytest ** 2))
print('\n\n','*'*80,'\nrun',run)
for num_examples in L_num_examples:
print('______\nsample size:', num_examples)
print('Current Experiment: Addition with noise '+str(noise_level) +' and ' + str(num_states)+' states')
data_function = lambda l: generate_data_simple_addition(num_examples, l, noise=noise_level)
Xl, yl = data_function(length)
X2l, y2l = data_function(length * 2)
X2l1, y2l1 = data_function(length * 2 + 1)
for method in methods:
if method != 'LSTM' and method != 'TIHT+SGD':
Tl = learning.sequence_to_tensor(Xl)
T2l = learning.sequence_to_tensor(X2l)
T2l1 = learning.sequence_to_tensor(X2l1)
t=tic()
Hl = learning.approximate_hankel(Tl, yl, alpha_ini_value=alpha,
rank=num_states, eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate, max_iters=TIHT_max_iters,
method=method, verbose=-1)
H2l = learning.approximate_hankel(T2l, y2l, alpha_ini_value=alpha,
rank=num_states, eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate, max_iters=TIHT_max_iters,
method=method, verbose=-1)
H2l1 = learning.approximate_hankel(T2l1, y2l1, alpha_ini_value=alpha, rank=num_states, eps=TIHT_epsilon,
learning_rate=TIHT_learning_rate, max_iters=TIHT_max_iters,
method=method, verbose=-1)
learned_model = learning.spectral_learning(num_states, H2l, H2l1, Hl)
test_mse = learning.compute_mse(learned_model, Xtest, ytest)
train_mse = learning.compute_mse(learned_model, X2l1, y2l1)
#print(test_mse)
if train_mse > np.mean(y2l1 ** 2):
test_mse = np.mean(ytest ** 2)
print(method,"test MSE:", test_mse, "\t\ttime:",toc(t))
results[method][num_examples].append(test_mse)
times[method][num_examples].append(toc(t))
elif method == 'LSTM':
def padding_function(x, desired_length):
if desired_length <= x.shape[1]:
return x
x = np.insert(x, x.shape[1], np.zeros((desired_length - x.shape[1], 1, x.shape[2])), axis=1)
return x
Xl_padded = padding_function(Xl, test_length)
X2l_padded = padding_function(X2l, test_length)
X2l1_padded = padding_function(X2l1, test_length)
#Xtest = padding_function(Xtest, test_length)
X = np.concatenate((Xl_padded, X2l_padded, X2l1_padded))
Y = np.concatenate((yl, y2l, y2l1))
t = tic()
learned_model = learning.RNN_LSTM(X, Y, test_length, num_states, noise_level, 'RandomRNN')
test_mse = learning.compute_mse(learned_model, Xtest, ytest, lstm = True)
train_mse = learning.compute_mse(learned_model, X2l1_padded, y2l1, lstm = True)
#if train_mse > np.mean(y2l1 ** 2):
# test_mse = np.mean(ytest ** 2)
print(method, "test MSE:", test_mse, "\t\ttime:", toc(t))
results[method][num_examples].append(test_mse)
times[method][num_examples].append(toc(t))
elif method == 'TIHT+SGD':
X = []
Y = []
for i in range(length * 2 + 2):
tempx, tempy = data_function(i)
X.append(tempx)
Y.append(tempy)
t = tic()
if noise_level == 0.:
TIHT_learning_rate = 0.000001
learned_model = learning.TIHT_SGD_torch(X, Y, num_states, length, verbose, TIHT_epsilon, TIHT_learning_rate,
TIHT_max_iters,
lr2, epo2, b2, tol, alpha=1., lifting=False)
test_mse = learning.compute_mse(learned_model, Xtest, ytest, if_tc = True)
train_mse = learning.compute_mse(learned_model, X2l1, y2l1, if_tc = True)
if train_mse > np.mean(y2l1 ** 2):
test_mse = np.mean(ytest ** 2)
print(method, "test MSE:", test_mse, "\t\ttime:", toc(t))
results[method][num_examples].append(test_mse)
times[method][num_examples].append(toc(t))
with open(xp_path + 'results_'+str(num_states)+'_states.pickle','wb') as f:
pickle.dump(results,f)
with open(xp_path + 'times'+str(num_states)+'_states.pickle', 'wb') as f:
pickle.dump(times, f)