This repository has been archived by the owner on Aug 5, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 492
/
lstm_layer.cpp
281 lines (242 loc) · 10.5 KB
/
lstm_layer.cpp
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
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
/*
All modification made by Intel Corporation: © 2016 Intel Corporation
All contributions by the University of California:
Copyright (c) 2014, 2015, The Regents of the University of California (Regents)
All rights reserved.
All other contributions:
Copyright (c) 2014, 2015, the respective contributors
All rights reserved.
For the list of contributors go to https://github.com/BVLC/caffe/blob/master/CONTRIBUTORS.md
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
* Neither the name of Intel Corporation nor the names of its contributors
may be used to endorse or promote products derived from this software
without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <string>
#include <vector>
#include "caffe/blob.hpp"
#include "caffe/common.hpp"
#include "caffe/filler.hpp"
#include "caffe/layer.hpp"
#include "caffe/layers/lstm_layer.hpp"
#include "caffe/util/math_functions.hpp"
namespace caffe {
template <typename Dtype>
void LSTMLayer<Dtype>::RecurrentInputBlobNames(vector<string>* names) const {
names->resize(2);
(*names)[0] = "h_0";
(*names)[1] = "c_0";
}
template <typename Dtype>
void LSTMLayer<Dtype>::RecurrentOutputBlobNames(vector<string>* names) const {
names->resize(2);
(*names)[0] = "h_" + format_int(this->T_);
(*names)[1] = "c_T";
}
template <typename Dtype>
void LSTMLayer<Dtype>::RecurrentInputShapes(vector<BlobShape>* shapes) const {
const int num_output = this->layer_param_.recurrent_param().num_output();
const int num_blobs = 2;
shapes->resize(num_blobs);
for (int i = 0; i < num_blobs; ++i) {
(*shapes)[i].Clear();
(*shapes)[i].add_dim(1); // a single timestep
(*shapes)[i].add_dim(this->N_);
(*shapes)[i].add_dim(num_output);
}
}
template <typename Dtype>
void LSTMLayer<Dtype>::OutputBlobNames(vector<string>* names) const {
names->resize(1);
(*names)[0] = "h";
}
template <typename Dtype>
void LSTMLayer<Dtype>::FillUnrolledNet(NetParameter* net_param) const {
const int num_output = this->layer_param_.recurrent_param().num_output();
CHECK_GT(num_output, 0) << "num_output must be positive";
const FillerParameter& weight_filler =
this->layer_param_.recurrent_param().weight_filler();
const FillerParameter& bias_filler =
this->layer_param_.recurrent_param().bias_filler();
// Add generic LayerParameter's (without bottoms/tops) of layer types we'll
// use to save redundant code.
LayerParameter hidden_param;
hidden_param.set_type("InnerProduct");
hidden_param.mutable_inner_product_param()->set_num_output(num_output * 4);
hidden_param.mutable_inner_product_param()->set_bias_term(false);
hidden_param.mutable_inner_product_param()->set_axis(2);
hidden_param.mutable_inner_product_param()->
mutable_weight_filler()->CopyFrom(weight_filler);
LayerParameter biased_hidden_param(hidden_param);
biased_hidden_param.mutable_inner_product_param()->set_bias_term(true);
biased_hidden_param.mutable_inner_product_param()->
mutable_bias_filler()->CopyFrom(bias_filler);
LayerParameter sum_param;
sum_param.set_type("Eltwise");
sum_param.mutable_eltwise_param()->set_operation(
EltwiseParameter_EltwiseOp_SUM);
LayerParameter scale_param;
scale_param.set_type("Scale");
scale_param.mutable_scale_param()->set_axis(0);
LayerParameter slice_param;
slice_param.set_type("Slice");
slice_param.mutable_slice_param()->set_axis(0);
LayerParameter split_param;
split_param.set_type("Split");
vector<BlobShape> input_shapes;
RecurrentInputShapes(&input_shapes);
CHECK_EQ(2, input_shapes.size());
LayerParameter* input_layer_param = net_param->add_layer();
input_layer_param->set_type("Input");
InputParameter* input_param = input_layer_param->mutable_input_param();
input_layer_param->add_top("c_0");
input_param->add_shape()->CopyFrom(input_shapes[0]);
input_layer_param->add_top("h_0");
input_param->add_shape()->CopyFrom(input_shapes[1]);
LayerParameter* cont_slice_param = net_param->add_layer();
cont_slice_param->CopyFrom(slice_param);
cont_slice_param->set_name("cont_slice");
cont_slice_param->add_bottom("cont");
cont_slice_param->mutable_slice_param()->set_axis(0);
// Add layer to transform all timesteps of x to the hidden state dimension.
// W_xc_x = W_xc * x + b_c
{
LayerParameter* x_transform_param = net_param->add_layer();
x_transform_param->CopyFrom(biased_hidden_param);
x_transform_param->set_name("x_transform");
x_transform_param->add_param()->set_name("W_xc");
x_transform_param->add_param()->set_name("b_c");
x_transform_param->add_bottom("x");
x_transform_param->add_top("W_xc_x");
x_transform_param->add_propagate_down(true);
}
if (this->static_input_) {
// Add layer to transform x_static to the gate dimension.
// W_xc_x_static = W_xc_static * x_static
LayerParameter* x_static_transform_param = net_param->add_layer();
x_static_transform_param->CopyFrom(hidden_param);
x_static_transform_param->mutable_inner_product_param()->set_axis(1);
x_static_transform_param->set_name("W_xc_x_static");
x_static_transform_param->add_param()->set_name("W_xc_static");
x_static_transform_param->add_bottom("x_static");
x_static_transform_param->add_top("W_xc_x_static_preshape");
x_static_transform_param->add_propagate_down(true);
LayerParameter* reshape_param = net_param->add_layer();
reshape_param->set_type("Reshape");
BlobShape* new_shape =
reshape_param->mutable_reshape_param()->mutable_shape();
new_shape->add_dim(1); // One timestep.
// Should infer this->N as the dimension so we can reshape on batch size.
new_shape->add_dim(-1);
new_shape->add_dim(
x_static_transform_param->inner_product_param().num_output());
reshape_param->set_name("W_xc_x_static_reshape");
reshape_param->add_bottom("W_xc_x_static_preshape");
reshape_param->add_top("W_xc_x_static");
}
LayerParameter* x_slice_param = net_param->add_layer();
x_slice_param->CopyFrom(slice_param);
x_slice_param->add_bottom("W_xc_x");
x_slice_param->set_name("W_xc_x_slice");
LayerParameter output_concat_layer;
output_concat_layer.set_name("h_concat");
output_concat_layer.set_type("Concat");
output_concat_layer.add_top("h");
output_concat_layer.mutable_concat_param()->set_axis(0);
for (int t = 1; t <= this->T_; ++t) {
string tm1s = format_int(t - 1);
string ts = format_int(t);
cont_slice_param->add_top("cont_" + ts);
x_slice_param->add_top("W_xc_x_" + ts);
// Add layers to flush the hidden state when beginning a new
// sequence, as indicated by cont_t.
// h_conted_{t-1} := cont_t * h_{t-1}
//
// Normally, cont_t is binary (i.e., 0 or 1), so:
// h_conted_{t-1} := h_{t-1} if cont_t == 1
// 0 otherwise
{
LayerParameter* cont_h_param = net_param->add_layer();
cont_h_param->CopyFrom(scale_param);
cont_h_param->set_name("h_conted_" + tm1s);
cont_h_param->add_bottom("h_" + tm1s);
cont_h_param->add_bottom("cont_" + ts);
cont_h_param->add_top("h_conted_" + tm1s);
}
// Add layer to compute
// W_hc_h_{t-1} := W_hc * h_conted_{t-1}
{
LayerParameter* w_param = net_param->add_layer();
w_param->CopyFrom(hidden_param);
w_param->set_name("transform_" + ts);
w_param->add_param()->set_name("W_hc");
w_param->add_bottom("h_conted_" + tm1s);
w_param->add_top("W_hc_h_" + tm1s);
w_param->mutable_inner_product_param()->set_axis(2);
}
// Add the outputs of the linear transformations to compute the gate input.
// gate_input_t := W_hc * h_conted_{t-1} + W_xc * x_t + b_c
// = W_hc_h_{t-1} + W_xc_x_t + b_c
{
LayerParameter* input_sum_layer = net_param->add_layer();
input_sum_layer->CopyFrom(sum_param);
input_sum_layer->set_name("gate_input_" + ts);
input_sum_layer->add_bottom("W_hc_h_" + tm1s);
input_sum_layer->add_bottom("W_xc_x_" + ts);
if (this->static_input_) {
input_sum_layer->add_bottom("W_xc_x_static");
}
input_sum_layer->add_top("gate_input_" + ts);
}
// Add LSTMUnit layer to compute the cell & hidden vectors c_t and h_t.
// Inputs: c_{t-1}, gate_input_t = (i_t, f_t, o_t, g_t), cont_t
// Outputs: c_t, h_t
// [ i_t' ]
// [ f_t' ] := gate_input_t
// [ o_t' ]
// [ g_t' ]
// i_t := \sigmoid[i_t']
// f_t := \sigmoid[f_t']
// o_t := \sigmoid[o_t']
// g_t := \tanh[g_t']
// c_t := cont_t * (f_t .* c_{t-1}) + (i_t .* g_t)
// h_t := o_t .* \tanh[c_t]
{
LayerParameter* lstm_unit_param = net_param->add_layer();
lstm_unit_param->set_type("LSTMUnit");
lstm_unit_param->add_bottom("c_" + tm1s);
lstm_unit_param->add_bottom("gate_input_" + ts);
lstm_unit_param->add_bottom("cont_" + ts);
lstm_unit_param->add_top("c_" + ts);
lstm_unit_param->add_top("h_" + ts);
lstm_unit_param->set_name("unit_" + ts);
}
output_concat_layer.add_bottom("h_" + ts);
} // for (int t = 1; t <= this->T_; ++t)
{
LayerParameter* c_T_copy_param = net_param->add_layer();
c_T_copy_param->CopyFrom(split_param);
c_T_copy_param->add_bottom("c_" + format_int(this->T_));
c_T_copy_param->add_top("c_T");
}
net_param->add_layer()->CopyFrom(output_concat_layer);
}
INSTANTIATE_CLASS(LSTMLayer);
REGISTER_LAYER_CLASS(LSTM);
} // namespace caffe