|
1
|
|
|
#!/usr/bin/env python |
|
2
|
|
|
# -*- coding: utf-8 -*- |
|
3
|
|
|
|
|
4
|
|
|
import theano.tensor as T |
|
5
|
|
|
from recurrent import RecurrentLayer |
|
6
|
|
|
|
|
7
|
|
|
OUTPUT_TYPES = ["sequence", "one"] |
|
8
|
|
|
INPUT_TYPES = ["sequence", "one"] |
|
9
|
|
|
|
|
10
|
|
|
class GRU(RecurrentLayer): |
|
11
|
|
|
|
|
12
|
|
|
def __init__(self, hidden_size, **kwargs): |
|
13
|
|
|
kwargs["hidden_size"] = hidden_size |
|
14
|
|
|
super(GRU, self).__init__("GRU", ["state"], **kwargs) |
|
15
|
|
|
|
|
16
|
|
|
def compute_new_state(self, step_inputs): |
|
17
|
|
|
xz_t, xr_t, xh_t, h_tm1 = map(step_inputs.get, ["xz_t", "xr_t", "xh_t", "state"]) |
|
18
|
|
|
if not xz_t: |
|
19
|
|
|
xz_t, xr_t, xh_t = 0, 0, 0 |
|
20
|
|
|
|
|
21
|
|
|
z_t = self.gate_activate(xz_t + T.dot(h_tm1, self.U_z) + self.b_z) |
|
22
|
|
|
r_t = self.gate_activate(xr_t + T.dot(h_tm1, self.U_r) + self.b_r) |
|
23
|
|
|
h_t_pre = self.activate(xh_t + T.dot(r_t * h_tm1, self.U_h) + self.b_h) |
|
24
|
|
|
h_t = z_t * h_tm1 + (1 - z_t) * h_t_pre |
|
25
|
|
|
|
|
26
|
|
|
return {"state": h_t} |
|
27
|
|
|
|
|
28
|
|
|
def merge_inputs(self, input_var, additional_inputs=None): |
|
29
|
|
|
if not additional_inputs: |
|
30
|
|
|
additional_inputs = [] |
|
31
|
|
|
all_inputs = [input_var] + additional_inputs |
|
32
|
|
|
z_inputs = [] |
|
33
|
|
|
r_inputs = [] |
|
34
|
|
|
h_inputs = [] |
|
35
|
|
|
for x, weights in zip(all_inputs, self.input_weights): |
|
36
|
|
|
wz, wr, wh = weights |
|
37
|
|
|
z_inputs.append(T.dot(x, wz)) |
|
38
|
|
|
r_inputs.append(T.dot(x, wr)) |
|
39
|
|
|
h_inputs.append(T.dot(x, wh)) |
|
40
|
|
|
merged_inputs = { |
|
41
|
|
|
"xz_t": sum(z_inputs), |
|
42
|
|
|
"xr_t": sum(r_inputs), |
|
43
|
|
|
"xh_t": sum(h_inputs) |
|
44
|
|
|
} |
|
45
|
|
|
return merged_inputs |
|
46
|
|
|
|
|
47
|
|
|
def prepare(self): |
|
48
|
|
|
self.output_dim = self.hidden_size |
|
49
|
|
|
|
|
50
|
|
|
self.U_z = self.create_weight(self.hidden_size, self.hidden_size, "uz", initializer=self.inner_init) |
|
51
|
|
|
self.b_z = self.create_bias(self.hidden_size, "z") |
|
52
|
|
|
|
|
53
|
|
|
self.U_r = self.create_weight(self.hidden_size, self.hidden_size, "ur", initializer=self.inner_init) |
|
54
|
|
|
self.b_r = self.create_bias(self.hidden_size, "r") |
|
55
|
|
|
|
|
56
|
|
|
self.U_h = self.create_weight(self.hidden_size, self.hidden_size, "uh", initializer=self.inner_init) |
|
57
|
|
|
self.b_h = self.create_bias(self.hidden_size, "h") |
|
58
|
|
|
|
|
59
|
|
|
self.register_parameters(self.U_z, self.b_z, |
|
60
|
|
|
self.U_r, self.b_r, |
|
61
|
|
|
self.U_h, self.b_h) |
|
62
|
|
|
|
|
63
|
|
|
self.input_weights = [] |
|
64
|
|
|
if self._input_type == "sequence": |
|
65
|
|
|
all_input_dims = [self.input_dim] + self.additional_input_dims |
|
66
|
|
|
for i, input_dim in enumerate(all_input_dims): |
|
67
|
|
|
wz = self.create_weight(input_dim, self.hidden_size, "wz_{}".format(i+1), initializer=self.outer_init) |
|
68
|
|
|
wr = self.create_weight(input_dim, self.hidden_size, "wr_{}".format(i+1), initializer=self.outer_init) |
|
69
|
|
|
wh = self.create_weight(input_dim, self.hidden_size, "wh_{}".format(i+1), initializer=self.outer_init) |
|
70
|
|
|
weights = [wz, wr, wh] |
|
71
|
|
|
self.input_weights.append(weights) |
|
72
|
|
|
self.register_parameters(*weights) |
|
73
|
|
|
|
uaca /
deepy
