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
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
|
#pragma once
#include "marian.h"
#include "layers/constructors.h"
#include "rnn/constructors.h"
#include "rnn/attention_constructors.h"
namespace marian {
class EncoderS2S : public EncoderBase {
public:
Expr applyEncoderRNN(Ptr<ExpressionGraph> graph,
Expr embeddings,
Expr mask,
std::string type) {
int first, second;
if(type == "bidirectional" || type == "alternating") {
// build two separate stacks, concatenate top output
first = opt<int>("enc-depth");
second = 0;
} else {
// build 1-layer bidirectional stack, concatenate,
// build n-1 layer unidirectional stack
first = 1;
second = opt<int>("enc-depth") - first;
}
auto forward = type == "alternating" ? rnn::dir::alternating_forward
: rnn::dir::forward;
auto backward = type == "alternating" ? rnn::dir::alternating_backward
: rnn::dir::backward;
using namespace keywords;
float dropoutRnn = inference_ ? 0 : opt<float>("dropout-rnn");
auto rnnFw = rnn::rnn(graph) //
("type", opt<std::string>("enc-cell")) //
("direction", (int)forward) //
("dimInput", embeddings->shape()[-1]) //
("dimState", opt<int>("dim-rnn")) //
("dropout", dropoutRnn) //
("layer-normalization", opt<bool>("layer-normalization")) //
("skip", opt<bool>("skip"));
for(int i = 1; i <= first; ++i) {
auto stacked = rnn::stacked_cell(graph);
for(int j = 1; j <= opt<int>("enc-cell-depth"); ++j) {
std::string paramPrefix = prefix_ + "_bi";
if(i > 1)
paramPrefix += "_l" + std::to_string(i);
if(i > 1 || j > 1)
paramPrefix += "_cell" + std::to_string(j);
bool transition = (j > 1);
stacked.push_back(rnn::cell(graph) //
("prefix", paramPrefix) //
("transition", transition));
}
rnnFw.push_back(stacked);
}
auto rnnBw = rnn::rnn(graph) //
("type", opt<std::string>("enc-cell")) //
("direction", (int)backward) //
("dimInput", embeddings->shape()[-1]) //
("dimState", opt<int>("dim-rnn")) //
("dropout", dropoutRnn) //
("layer-normalization", opt<bool>("layer-normalization")) //
("skip", opt<bool>("skip"));
for(int i = 1; i <= first; ++i) {
auto stacked = rnn::stacked_cell(graph);
for(int j = 1; j <= opt<int>("enc-cell-depth"); ++j) {
std::string paramPrefix = prefix_ + "_bi_r";
if(i > 1)
paramPrefix += "_l" + std::to_string(i);
if(i > 1 || j > 1)
paramPrefix += "_cell" + std::to_string(j);
bool transition = (j > 1);
stacked.push_back(rnn::cell(graph) //
("prefix", paramPrefix) //
("transition", transition));
}
rnnBw.push_back(stacked);
}
auto context = concatenate({rnnFw->transduce(embeddings, mask),
rnnBw->transduce(embeddings, mask)},
axis = -1);
if(second > 0) {
// add more layers (unidirectional) by transducing the output of the
// previous bidirectional RNN through multiple layers
// construct RNN first
auto rnnUni = rnn::rnn(graph) //
("type", opt<std::string>("enc-cell")) //
("dimInput", 2 * opt<int>("dim-rnn")) //
("dimState", opt<int>("dim-rnn")) //
("dropout", dropoutRnn) //
("layer-normalization", opt<bool>("layer-normalization")) //
("skip", opt<bool>("skip"));
for(int i = first + 1; i <= second + first; ++i) {
auto stacked = rnn::stacked_cell(graph);
for(int j = 1; j <= opt<int>("enc-cell-depth"); ++j) {
std::string paramPrefix = prefix_ + "_l" + std::to_string(i) + "_cell"
+ std::to_string(j);
stacked.push_back(rnn::cell(graph)("prefix", paramPrefix));
}
rnnUni.push_back(stacked);
}
// transduce context to new context
context = rnnUni->transduce(context);
}
return context;
}
Expr buildSourceEmbeddings(Ptr<ExpressionGraph> graph) {
// create source embeddings
int dimVoc = opt<std::vector<int>>("dim-vocabs")[batchIndex_];
int dimEmb = opt<int>("dim-emb");
auto embFactory = embedding(graph) //
("dimVocab", dimVoc) //
("dimEmb", dimEmb);
if(opt<bool>("tied-embeddings-src") || opt<bool>("tied-embeddings-all"))
embFactory("prefix", "Wemb");
else
embFactory("prefix", prefix_ + "_Wemb");
if(options_->has("embedding-fix-src"))
embFactory("fixed", opt<bool>("embedding-fix-src"));
if(options_->has("embedding-vectors")) {
auto embFiles = opt<std::vector<std::string>>("embedding-vectors");
embFactory //
("embFile", embFiles[batchIndex_]) //
("normalization", opt<bool>("embedding-normalization"));
}
return embFactory.construct();
}
EncoderS2S(Ptr<Options> options) : EncoderBase(options) {}
virtual Ptr<EncoderState> build(Ptr<ExpressionGraph> graph,
Ptr<data::CorpusBatch> batch) {
auto embeddings = buildSourceEmbeddings(graph);
using namespace keywords;
// select embeddings that occur in the batch
Expr batchEmbeddings, batchMask;
std::tie(batchEmbeddings, batchMask)
= EncoderBase::lookup(graph, embeddings, batch);
// apply dropout over source words
float dropProb = inference_ ? 0 : opt<float>("dropout-src");
if(dropProb) {
int srcWords = batchEmbeddings->shape()[-3];
batchEmbeddings = dropout(batchEmbeddings, dropProb, {srcWords, 1, 1});
}
Expr context = applyEncoderRNN(
graph, batchEmbeddings, batchMask, opt<std::string>("enc-type"));
return New<EncoderState>(context, batchMask, batch);
}
void clear() {}
};
class DecoderS2S : public DecoderBase {
private:
Ptr<rnn::RNN> rnn_;
Ptr<rnn::RNN> constructDecoderRNN(Ptr<ExpressionGraph> graph,
Ptr<DecoderState> state) {
float dropoutRnn = inference_ ? 0 : opt<float>("dropout-rnn");
auto rnn = rnn::rnn(graph) //
("type", opt<std::string>("dec-cell")) //
("dimInput", opt<int>("dim-emb")) //
("dimState", opt<int>("dim-rnn")) //
("dropout", dropoutRnn) //
("layer-normalization", opt<bool>("layer-normalization")) //
("nematus-normalization",
options_->has("original-type")
&& opt<std::string>("original-type") == "nematus") //
("skip", opt<bool>("skip"));
size_t decoderLayers = opt<size_t>("dec-depth");
size_t decoderBaseDepth = opt<size_t>("dec-cell-base-depth");
size_t decoderHighDepth = opt<size_t>("dec-cell-high-depth");
// setting up conditional (transitional) cell
auto baseCell = rnn::stacked_cell(graph);
for(int i = 1; i <= decoderBaseDepth; ++i) {
bool transition = (i > 2);
auto paramPrefix = prefix_ + "_cell" + std::to_string(i);
baseCell.push_back(rnn::cell(graph) //
("prefix", paramPrefix) //
("final", i > 1) //
("transition", transition));
if(i == 1) {
for(int k = 0; k < state->getEncoderStates().size(); ++k) {
auto attPrefix = prefix_;
if(state->getEncoderStates().size() > 1)
attPrefix += "_att" + std::to_string(k + 1);
auto encState = state->getEncoderStates()[k];
baseCell.push_back(
rnn::attention(graph)("prefix", attPrefix).set_state(encState));
}
}
}
// Add cell to RNN (first layer)
rnn.push_back(baseCell);
// Add more cells to RNN (stacked RNN)
for(int i = 2; i <= decoderLayers; ++i) {
// deep transition
auto highCell = rnn::stacked_cell(graph);
for(int j = 1; j <= decoderHighDepth; j++) {
auto paramPrefix
= prefix_ + "_l" + std::to_string(i) + "_cell" + std::to_string(j);
highCell.push_back(rnn::cell(graph)("prefix", paramPrefix));
}
// Add cell to RNN (more layers)
rnn.push_back(highCell);
}
return rnn.construct();
}
public:
DecoderS2S(Ptr<Options> options) : DecoderBase(options) {}
virtual Ptr<DecoderState> startState(
Ptr<ExpressionGraph> graph,
Ptr<data::CorpusBatch> batch,
std::vector<Ptr<EncoderState>>& encStates) {
using namespace keywords;
std::vector<Expr> meanContexts;
for(auto& encState : encStates) {
// average the source context weighted by the batch mask
// this will remove padded zeros from the average
meanContexts.push_back(weighted_average(
encState->getContext(), encState->getMask(), axis = -3));
}
Expr start;
if(!meanContexts.empty()) {
// apply single layer network to mean to map into decoder space
auto mlp = mlp::mlp(graph).push_back(
mlp::dense(graph) //
("prefix", prefix_ + "_ff_state") //
("dim", opt<int>("dim-rnn")) //
("activation", (int)mlp::act::tanh) //
("layer-normalization", opt<bool>("layer-normalization")) //
("nematus-normalization",
options_->has("original-type")
&& opt<std::string>("original-type") == "nematus") //
);
start = mlp->apply(meanContexts);
} else {
int dimBatch = batch->size();
int dimRnn = opt<int>("dim-rnn");
start = graph->constant({dimBatch, dimRnn}, inits::zeros);
}
rnn::States startStates(opt<size_t>("dec-depth"), {start, start});
return New<DecoderState>(startStates, nullptr, encStates);
}
virtual Ptr<DecoderState> step(Ptr<ExpressionGraph> graph,
Ptr<DecoderState> state) {
using namespace keywords;
auto embeddings = state->getTargetEmbeddings();
// dropout target words
float dropoutTrg = inference_ ? 0 : opt<float>("dropout-trg");
if(dropoutTrg) {
int trgWords = embeddings->shape()[-3];
embeddings = dropout(embeddings, dropoutTrg, {trgWords, 1, 1});
}
if(!rnn_)
rnn_ = constructDecoderRNN(graph, state);
// apply RNN to embeddings, initialized with encoder context mapped into
// decoder space
auto decoderContext = rnn_->transduce(embeddings, state->getStates());
// retrieve the last state per layer. They are required during translation
// in order to continue decoding for the next word
rnn::States decoderStates = rnn_->lastCellStates();
std::vector<Expr> alignedContexts;
for(int k = 0; k < state->getEncoderStates().size(); ++k) {
// retrieve all the aligned contexts computed by the attention mechanism
auto att = rnn_->at(0)
->as<rnn::StackedCell>()
->at(k + 1)
->as<rnn::Attention>();
alignedContexts.push_back(att->getContext());
}
Expr alignedContext;
if(alignedContexts.size() > 1)
alignedContext = concatenate(alignedContexts, axis = -1);
else if(alignedContexts.size() == 1)
alignedContext = alignedContexts[0];
// construct deep output multi-layer network layer-wise
auto layer1 = mlp::dense(graph) //
("prefix", prefix_ + "_ff_logit_l1") //
("dim", opt<int>("dim-emb")) //
("activation", (int)mlp::act::tanh) //
("layer-normalization", opt<bool>("layer-normalization")) //
("nematus-normalization",
options_->has("original-type")
&& opt<std::string>("original-type") == "nematus");
int dimTrgVoc = opt<std::vector<int>>("dim-vocabs")[batchIndex_];
auto layer2 = mlp::dense(graph) //
("prefix", prefix_ + "_ff_logit_l2") //
("dim", dimTrgVoc);
if(opt<bool>("tied-embeddings") || opt<bool>("tied-embeddings-all")) {
std::string tiedPrefix = prefix_ + "_Wemb";
if(opt<bool>("tied-embeddings-all") || opt<bool>("tied-embeddings-src"))
tiedPrefix = "Wemb";
layer2.tie_transposed("W", tiedPrefix);
}
// assemble layers into MLP and apply to embeddings, decoder context and
// aligned source context
auto output = mlp::mlp(graph) //
.push_back(layer1) //
.push_back(layer2);
Expr logits;
if(alignedContext)
logits = output->apply(embeddings, decoderContext, alignedContext);
else
logits = output->apply(embeddings, decoderContext);
// return unormalized(!) probabilities
return New<DecoderState>(decoderStates, logits, state->getEncoderStates());
}
// helper function for guided alignment
virtual const std::vector<Expr> getAlignments(int i = 0) {
auto att
= rnn_->at(0)->as<rnn::StackedCell>()->at(i + 1)->as<rnn::Attention>();
return att->getAlignments();
}
void clear() { rnn_ = nullptr; }
};
}
|