forked from gwtaylor/theano-rnn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
hf_example.py
225 lines (183 loc) · 7.47 KB
/
hf_example.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
"""
This code uses the recurrent neural net implementation in rnn.py
but trains it using Hessian-Free optimization.
It requires the theano-hf package:
https://github.com/boulanni/theano-hf
@author Graham Taylor
"""
from rnn import MetaRNN
from hf import SequenceDataset, hf_optimizer
import numpy as np
import matplotlib.pyplot as plt
import logging
def test_real(n_updates=100):
""" Test RNN with real-valued outputs. """
n_hidden = 10
n_in = 5
n_out = 3
n_steps = 10
n_seq = 1000
np.random.seed(0)
# simple lag test
seq = np.random.randn(n_seq, n_steps, n_in)
print 'seq1'
print seq
targets = np.zeros((n_seq, n_steps, n_out))
targets[:, 1:, 0] = seq[:, :-1, 3] # delayed 1
targets[:, 1:, 1] = seq[:, :-1, 2] # delayed 1
targets[:, 2:, 2] = seq[:, :-2, 0] # delayed 2
targets += 0.01 * np.random.standard_normal(targets.shape)
print 'targets'
print targets
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
print 'seq2'
print seq
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=100)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=20)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh')
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y)], h=model.rnn.h)
print model
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
plt.close('all')
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[0])
ax1.set_title('input')
ax2 = plt.subplot(212)
true_targets = plt.plot(targets[0])
guess = model.predict(seq[0])
guessed_targets = plt.plot(guess, linestyle='--')
for i, x in enumerate(guessed_targets):
x.set_color(true_targets[i].get_color())
ax2.set_title('solid: true output, dashed: model output')
def test_binary(multiple_out=False, n_updates=250):
""" Test RNN with binary outputs. """
n_hidden = 10
n_in = 5
if multiple_out:
n_out = 2
else:
n_out = 1
n_steps = 10
n_seq = 100
np.random.seed(0)
# simple lag test
seq = np.random.randn(n_seq, n_steps, n_in)
targets = np.zeros((n_seq, n_steps, n_out), dtype='int32')
# whether lag 1 (dim 3) is greater than lag 2 (dim 0)
targets[:, 2:, 0] = np.cast[np.int32](seq[:, 1:-1, 3] > seq[:, :-2, 0])
if multiple_out:
# whether product of lag 1 (dim 4) and lag 1 (dim 2)
# is less than lag 2 (dim 0)
targets[:, 2:, 1] = np.cast[np.int32](
(seq[:, 1:-1, 4] * seq[:, 1:-1, 2]) > seq[:, :-2, 0])
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh', output_type='binary')
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
# using settings of initial_lambda and mu given in Nicolas' RNN example
# seem to do a little worse than the default
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
seqs = xrange(10)
plt.close('all')
for seq_num in seqs:
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[seq_num])
ax1.set_title('input')
ax2 = plt.subplot(212)
true_targets = plt.step(xrange(n_steps), targets[seq_num], marker='o')
guess = model.predict_proba(seq[seq_num])
guessed_targets = plt.step(xrange(n_steps), guess)
plt.setp(guessed_targets, linestyle='--', marker='d')
for i, x in enumerate(guessed_targets):
x.set_color(true_targets[i].get_color())
ax2.set_ylim((-0.1, 1.1))
ax2.set_title('solid: true output, dashed: model output (prob)')
def test_softmax(n_updates=250):
""" Test RNN with softmax outputs. """
n_hidden = 10
n_in = 5
n_steps = 10
n_seq = 100
n_classes = 3
n_out = n_classes # restricted to single softmax per time step
np.random.seed(0)
# simple lag test
seq = np.random.randn(n_seq, n_steps, n_in)
targets = np.zeros((n_seq, n_steps), dtype='int32')
thresh = 0.5
# if lag 1 (dim 3) is greater than lag 2 (dim 0) + thresh
# class 1
# if lag 1 (dim 3) is less than lag 2 (dim 0) - thresh
# class 2
# if lag 2(dim0) - thresh <= lag 1 (dim 3) <= lag2(dim0) + thresh
# class 0
targets[:, 2:][seq[:, 1:-1, 3] > seq[:, :-2, 0] + thresh] = 1
targets[:, 2:][seq[:, 1:-1, 3] < seq[:, :-2, 0] - thresh] = 2
#targets[:, 2:, 0] = np.cast[np.int](seq[:, 1:-1, 3] > seq[:, :-2, 0])
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh', output_type='softmax',
use_symbolic_softmax=True)
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
# using settings of initial_lambda and mu given in Nicolas' RNN example
# seem to do a little worse than the default
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
seqs = xrange(10)
plt.close('all')
for seq_num in seqs:
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[seq_num])
ax1.set_title('input')
ax2 = plt.subplot(212)
# blue line will represent true classes
true_targets = plt.step(xrange(n_steps), targets[seq_num], marker='o')
# show probabilities (in b/w) output by model
guess = model.predict_proba(seq[seq_num])
guessed_probs = plt.imshow(guess.T, interpolation='nearest',
cmap='gray')
ax2.set_title('blue: true class, grayscale: probs assigned by model')
def run():
logging.basicConfig(level=logging.INFO)
#test_real(n_updates=20)
test_binary(multiple_out=True, n_updates=20)
#test_softmax(n_updates=20)
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO)
test_real(n_updates=20)
#test_binary(multiple_out=True, n_updates=20)
#test_softmax(n_updates=20)