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
|
#pragma once
#include "layers/generic.h"
#include "marian.h"
#include "rnn/cells.h"
#include "rnn/types.h"
#include <algorithm>
#include <chrono>
#include <cstdio>
#include <iomanip>
#include <string>
namespace marian {
namespace rnn {
enum struct dir : int {
forward,
backward,
alternating_forward,
alternating_backward
};
}
} // namespace marian
YAML_REGISTER_TYPE(marian::rnn::dir, int)
namespace marian {
namespace rnn {
class BaseRNN {
protected:
Ptr<ExpressionGraph> graph_;
Ptr<Options> options_;
public:
BaseRNN(Ptr<ExpressionGraph> graph, Ptr<Options> options)
: graph_(graph), options_(options) {}
virtual Expr transduce(Expr, Expr = nullptr) = 0;
virtual Expr transduce(Expr, State, Expr = nullptr) = 0;
virtual Expr transduce(Expr, States, Expr = nullptr) = 0;
virtual States lastCellStates() = 0;
virtual void push_back(Ptr<Cell>) = 0;
virtual Ptr<Cell> at(int i) = 0;
virtual Ptr<Options> getOptions() { return options_; }
};
class RNN;
class SingleLayerRNN : public BaseRNN {
private:
Ptr<Cell> cell_;
dir direction_;
States last_;
States apply(const Expr input,
const States initialState,
const Expr mask = nullptr) {
last_.clear();
State state = initialState.front();
cell_->clear();
auto xWs = cell_->applyInput({input});
auto timeSteps = input->shape()[-3];
States outputs;
for(int i = 0; i < timeSteps; ++i) {
int j = i;
if(direction_ == dir::backward)
j = timeSteps - i - 1;
std::vector<Expr> steps(xWs.size());
std::transform(xWs.begin(), xWs.end(), steps.begin(), [j](Expr e) {
return step(e, j, -3);
});
if(mask)
state = cell_->applyState(steps, state, step(mask, j, -3));
else
state = cell_->applyState(steps, state);
outputs.push_back(state);
}
if(direction_ == dir::backward)
outputs.reverse();
last_.push_back(outputs.back());
return outputs;
}
States apply(const Expr input, const Expr mask = nullptr) {
auto graph = input->graph();
int dimBatch = input->shape()[-2];
int dimState = cell_->getOptions()->get<int>("dimState");
auto output = graph->zeros({1, dimBatch, dimState});
Expr cell = output;
State startState{output, cell};
return apply(input, States({startState}), mask);
}
SingleLayerRNN(Ptr<ExpressionGraph> graph, Ptr<Options> options)
: BaseRNN(graph, options),
direction_((dir)options->get<int>("direction", (int)dir::forward)) {}
public:
friend RNN;
// @TODO: benchmark whether this concatenation is a good idea
virtual Expr transduce(Expr input, Expr mask = nullptr) override {
return apply(input, mask).outputs();
}
virtual Expr transduce(Expr input, States states, Expr mask = nullptr) override {
return apply(input, states, mask).outputs();
}
virtual Expr transduce(Expr input, State state, Expr mask = nullptr) override {
return apply(input, States({state}), mask).outputs();
}
States lastCellStates() override { return last_; }
void push_back(Ptr<Cell> cell) override { cell_ = cell; }
virtual Ptr<Cell> at(int i) override {
ABORT_IF(i > 0, "SingleRNN only has one cell");
return cell_;
}
};
class RNN : public BaseRNN, public std::enable_shared_from_this<RNN> {
private:
bool skip_;
bool skipFirst_;
std::vector<Ptr<SingleLayerRNN>> rnns_;
public:
RNN(Ptr<ExpressionGraph> graph, Ptr<Options> options)
: BaseRNN(graph, options),
skip_(options->get("skip", false)),
skipFirst_(options->get("skipFirst", false)) {}
void push_back(Ptr<Cell> cell) override {
auto rnn
= Ptr<SingleLayerRNN>(new SingleLayerRNN(graph_, cell->getOptions()));
rnn->push_back(cell);
rnns_.push_back(rnn);
}
Expr transduce(Expr input, Expr mask = nullptr) override {
ABORT_IF(rnns_.empty(), "0 layers in RNN");
Expr output;
Expr layerInput = input;
for(size_t i = 0; i < rnns_.size(); ++i) {
auto lazyInput = layerInput;
auto cell = rnns_[i]->at(0);
auto lazyInputs = cell->getLazyInputs(shared_from_this());
if(!lazyInputs.empty()) {
lazyInputs.push_back(layerInput);
lazyInput = concatenate(lazyInputs, /*axis =*/ -1);
}
auto layerOutput = rnns_[i]->transduce(lazyInput, mask);
if(skip_ && (skipFirst_ || i > 0))
output = layerOutput + layerInput;
else
output = layerOutput;
layerInput = output;
}
return output;
}
Expr transduce(Expr input, States states, Expr mask = nullptr) override {
ABORT_IF(rnns_.empty(), "0 layers in RNN");
Expr output;
Expr layerInput = input;
for(size_t i = 0; i < rnns_.size(); ++i) {
Expr lazyInput;
auto cell = rnns_[i]->at(0);
auto lazyInputs = cell->getLazyInputs(shared_from_this());
if(!lazyInputs.empty()) {
lazyInputs.push_back(layerInput);
lazyInput = concatenate(lazyInputs, /*axis =*/ -1);
} else {
lazyInput = layerInput;
}
auto layerOutput
= rnns_[i]->transduce(lazyInput, States({states[i]}), mask);
if(skip_ && (skipFirst_ || i > 0))
output = layerOutput + layerInput;
else
output = layerOutput;
layerInput = output;
}
return output;
}
Expr transduce(Expr input, State state, Expr mask = nullptr) override {
ABORT_IF(rnns_.empty(), "0 layers in RNN");
Expr output;
Expr layerInput = input;
for(size_t i = 0; i < rnns_.size(); ++i) {
auto lazyInput = layerInput;
auto cell = rnns_[i]->at(0);
auto lazyInputs = cell->getLazyInputs(shared_from_this());
if(!lazyInputs.empty()) {
lazyInputs.push_back(layerInput);
lazyInput = concatenate(lazyInputs, /*axis =*/ -1);
}
auto layerOutput = rnns_[i]->transduce(lazyInput, States({state}), mask);
if(skip_ && (skipFirst_ || i > 0))
output = layerOutput + layerInput;
else
output = layerOutput;
layerInput = output;
}
return output;
}
States lastCellStates() override {
States temp;
for(auto rnn : rnns_)
temp.push_back(rnn->lastCellStates().back());
return temp;
}
virtual Ptr<Cell> at(int i) override { return rnns_[i]->at(0); }
};
} // namespace rnn
} // namespace marian
|
