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
|
import theano
from theano import tensor
import numpy
from blocks.algorithms import Momentum, AdaDelta, RMSProp
from blocks.bricks import Tanh, Softmax, Linear, MLP
from blocks.bricks.recurrent import LSTM
from blocks.initialization import IsotropicGaussian, Constant
from blocks.filter import VariableFilter
from blocks.roles import WEIGHT
from blocks.graph import ComputationGraph, apply_noise, apply_dropout
# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
# divided in chunks of lengh 'seq_div_size'
num_seqs = 20
seq_len = 2000
seq_div_size = 100
io_dim = 256
hidden_dims = [512, 512, 512]
activation_function = Tanh()
i2h_all = True # input to all hidden layers or only first layer
h2o_all = True # all hiden layers to output or only last layer
w_noise_std = 0.01
i_dropout = 0.5
step_rule = 'momentum'
learning_rate = 0.1
momentum = 0.9
param_desc = '%s-%sIH,%sHO-n%s-d%s-%dx%d(%d)-%s' % (
repr(hidden_dims),
'all' if i2h_all else 'first',
'all' if h2o_all else 'last',
repr(w_noise_std),
repr(i_dropout),
num_seqs, seq_len, seq_div_size,
step_rule
)
save_freq = 5
# parameters for sample generation
sample_len = 60
sample_temperature = 0.3
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)
else:
assert(False)
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'
# Construct hidden states
dims = [io_dim] + hidden_dims
states = [in_onehot.dimshuffle(1, 0, 2)]
bricks = []
updates = []
for i in xrange(1, len(dims)):
init_state = theano.shared(numpy.zeros((num_seqs, dims[i])).astype(theano.config.floatX),
name='st0_%d'%i)
init_cell = theano.shared(numpy.zeros((num_seqs, dims[i])).astype(theano.config.floatX),
name='cell0_%d'%i)
linear = Linear(input_dim=dims[i-1], output_dim=4*dims[i],
name="lstm_in_%d"%i)
bricks.append(linear)
inter = linear.apply(states[-1])
if i2h_all and i > 1:
linear2 = Linear(input_dim=dims[0], output_dim=4*dims[i],
name="lstm_in0_%d"%i)
bricks.append(linear2)
inter = inter + linear2.apply(states[0])
inter.name = 'inter_bis_%d'%i
lstm = LSTM(dim=dims[i], activation=activation_function,
name="lstm_rec_%d"%i)
bricks.append(lstm)
new_states, new_cells = lstm.apply(inter,
states=init_state,
cells=init_cell)
updates.append((init_state, new_states[-1, :, :]))
updates.append((init_cell, new_cells[-1, :, :]))
states.append(new_states)
states = [s.dimshuffle(1, 0, 2) for s in states]
# Construct output from hidden states
out = None
layers = zip(dims, states)[1:]
if not h2o_all:
layers = [layers[-1]]
for i, (dim, state) in enumerate(layers):
top_linear = Linear(input_dim=dim, output_dim=io_dim,
name='top_linear_%d'%i)
bricks.append(top_linear)
out_i = top_linear.apply(state)
out = out_i if out is None else out + out_i
out.name = 'out_part_%d'%i
# 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 bricks:
brick.weights_init = IsotropicGaussian(0.1)
brick.biases_init = Constant(0.)
brick.initialize()
# Apply noise and dropout
cg = ComputationGraph([cost, error_rate])
if w_noise_std > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, w_noise_std)
if i_dropout > 0:
cg = apply_dropout(cg, states[1:], i_dropout)
[cost_reg, error_rate_reg] = cg.outputs
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
self.updates = updates
|
