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
|
import theano
from theano import tensor
import numpy
from blocks.algorithms import Momentum, AdaDelta, RMSProp, Adam
from blocks.bricks import Activation, Tanh, Logistic, Softmax, Rectifier, Linear, MLP, Initializable, Identity
from blocks.bricks.base import application, lazy
from blocks.bricks.recurrent import BaseRecurrent, recurrent
from blocks.initialization import IsotropicGaussian, Constant
from blocks.utils import shared_floatx_zeros
from blocks.filter import VariableFilter
from blocks.roles import WEIGHT, INITIAL_STATE, add_role
from blocks.graph import ComputationGraph, apply_noise, apply_dropout
class TRectifier(Activation):
@application(inputs=['input_'], outputs=['output'])
def apply(self, input_):
return tensor.switch(input_ > 1, input_, 0)
# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
# divided in chunks of lengh 'seq_div_size'
num_seqs = 10
seq_len = 2000
seq_div_size = 100
io_dim = 256
state_dim = 1024
activation = Tanh()
transition_hidden = [1024, 1024]
transition_hidden_activations = [Rectifier(), Rectifier()]
output_hidden = []
output_hidden_activations = []
weight_noise_std = 0.05
output_h_dropout = 0.5
l1_state = 0.01
l1_reset = 0.01
step_rule = 'momentum'
learning_rate = 0.01
momentum = 0.99
param_desc = '%s,t%s,o%s-n%s-d%s-L1:%s,%s-%s' % (
repr(state_dim), repr(transition_hidden), repr(output_hidden),
repr(weight_noise_std),
repr(output_h_dropout),
repr(l1_state), repr(l1_reset),
step_rule
)
save_freq = 5
on_irc = False
# parameters for sample generation
sample_len = 100
sample_temperature = 0.7 #0.5
sample_freq = 1
if step_rule == 'rmsprop':
step_rule = RMSProp()
elif step_rule == 'adadelta':
step_rule = AdaDelta()
elif step_rule == 'momentum':
step_rule = Momentum(learning_rate=learning_rate, momentum=momentum)
elif step_rule == 'adam':
step_rule = Adam()
else:
assert(False)
class DGSRNN(BaseRecurrent, Initializable):
def __init__(self, input_dim, state_dim, act, transition_h, tr_h_activations, **kwargs):
super(DGSRNN, self).__init__(**kwargs)
self.input_dim = input_dim
self.state_dim = state_dim
logistic = Logistic()
self.inter = MLP(dims=[input_dim + state_dim] + transition_h,
activations=tr_h_activations,
name='inter')
self.reset = MLP(dims=[transition_h[-1], state_dim],
activations=[logistic],
name='reset')
self.update = MLP(dims=[transition_h[-1], state_dim],
activations=[act],
name='update')
self.children = [self.inter, self.reset, self.update, logistic, act] + tr_h_activations
# init state
self.params = [shared_floatx_zeros((state_dim,), name='init_state')]
add_role(self.params[0], INITIAL_STATE)
def get_dim(self, name):
if name == 'state':
return self.state_dim
return super(GFGRU, self).get_dim(name)
@recurrent(sequences=['inputs'], states=['state'],
outputs=['state', 'reset'], contexts=[])
def apply(self, inputs=None, state=None):
inter_v = self.inter.apply(tensor.concatenate([inputs, state], axis=1))
reset_v = self.reset.apply(inter_v)
update_v = self.update.apply(inter_v)
new_state = state * (1 - reset_v) + reset_v * update_v
return new_state, reset_v
@application
def initial_state(self, state_name, batch_size, *args, **kwargs):
return tensor.repeat(self.params[0][None, :],
repeats=batch_size,
axis=0)
class Model():
def __init__(self):
inp = tensor.lmatrix('bytes')
in_onehot = tensor.eq(tensor.arange(io_dim, dtype='int16').reshape((1, 1, io_dim)),
inp[:, :, None])
in_onehot.name = 'in_onehot'
dgsrnn = DGSRNN(input_dim=io_dim,
state_dim=state_dim,
act=activation,
transition_h=transition_hidden,
tr_h_activations=transition_hidden_activations,
name='dgsrnn')
prev_state = theano.shared(numpy.zeros((num_seqs, state_dim)).astype(theano.config.floatX),
name='state')
states, resets = dgsrnn.apply(in_onehot.dimshuffle(1, 0, 2), state=prev_state)
states = states.dimshuffle(1, 0, 2)
resets = resets.dimshuffle(1, 0, 2)
self.states = [(prev_state, states[:, -1, :])]
out_mlp = MLP(dims=[state_dim] + output_hidden + [io_dim],
activations=output_hidden_activations + [None],
name='output_mlp')
states_sh = states.reshape((inp.shape[0]*inp.shape[1], state_dim))
out = out_mlp.apply(states_sh).reshape((inp.shape[0], inp.shape[1], io_dim))
# Do prediction and calculate cost
pred = out.argmax(axis=2)
cost = Softmax().categorical_cross_entropy(inp[:, 1:].flatten(),
out[:, :-1, :].reshape((inp.shape[0]*(inp.shape[1]-1),
io_dim)))
error_rate = tensor.neq(inp[:, 1:].flatten(), pred[:, :-1].flatten()).mean()
# Initialize all bricks
for brick in [dgsrnn, out_mlp]:
brick.weights_init = IsotropicGaussian(0.001)
brick.biases_init = Constant(0.0)
brick.initialize()
# Apply noise and dropout
cg = ComputationGraph([cost, error_rate, states, resets])
if weight_noise_std > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, weight_noise_std)
if output_h_dropout > 0:
dv = VariableFilter(name='input_', bricks=out_mlp.linear_transformations)(cg)
print "Output H dropout on", len(dv), "vars"
cg = apply_dropout(cg, dv, output_h_dropout)
[cost_reg, error_rate_reg, states, resets] = cg.outputs
if l1_state > 0:
cost_reg = cost_reg + l1_state * abs(states).mean()
if l1_reset > 0:
cost_reg = cost_reg + l1_reset * abs(resets).mean()
self.cost = cost
self.error_rate = error_rate
self.cost_reg = cost_reg
self.error_rate_reg = error_rate_reg
self.out = out
self.pred = pred
|
