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
|
#include "model.h"
#include "gpu/mblas/tensor_functions.h"
using namespace std;
namespace {
using namespace amunmt;
using namespace GPU;
using namespace mblas;
shared_ptr<Tensor> merge (shared_ptr<Tensor> m1, shared_ptr<Tensor> m2) {
if (m2->size()) {
Transpose(*m1);
Transpose(*m2);
Concat(*m1, *m2);
Transpose(*m1);
Transpose(*m2);
}
return m1;
}
}
namespace amunmt {
namespace GPU {
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::EncEmbeddings::EncEmbeddings(const NpzConverter& model) {
Es_.emplace_back(model.get("Wemb", true));
for(int i=1; true; i++) {
std::string factorKey = "Wemb" + std::to_string(i);
std::shared_ptr<mblas::Tensor> factorEmb = model.get(factorKey, false);
if (factorEmb->size() <= 0){
break;
}
Es_.emplace_back(factorEmb);
}
}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::EncForwardGRU::EncForwardGRU(const NpzConverter& model)
: W_(model.get("encoder_W", true)),
B_(model.get("encoder_b", true, true)),
U_(model.get("encoder_U", true)),
Wx_(model.get("encoder_Wx", true)),
Bx1_(model.get("encoder_bx", true, true)),
Bx2_(new mblas::Tensor(Bx1_->dim(0), Bx1_->dim(1), Bx1_->dim(2), Bx1_->dim(3), true)),
Ux_(model.get("encoder_Ux", true)),
Gamma_1_(model.get("encoder_gamma1", false)),
Gamma_2_(model.get("encoder_gamma2", false))
{ }
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::EncForwardLSTM::EncForwardLSTM(const NpzConverter& model)
// matrix merging is done to be backwards-compatible with the original LSTM implementation in Amun
// we now use the same format used in Marian
// TODO: adapt to support Nematus LSTM models which use a similar format to Amun's original format
: W_(merge(model.get("encoder_W", true), model.get("encoder_Wx", false))),
B_(merge(model.get("encoder_b", true, true), model.get("encoder_bx", false, true))),
U_(merge(model.get("encoder_U", true), model.get("encoder_Ux", false))),
Gamma_1_(model.get("encoder_gamma1", false)),
Gamma_2_(model.get("encoder_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::EncBackwardLSTM::EncBackwardLSTM(const NpzConverter& model)
// matrix merging is done to be backwards-compatible with the original LSTM implementation in Amun
// we now use the same format used in Marian
// TODO: adapt to support Nematus LSTM models which use a similar format to Amun's original format
: W_(merge(model.get("encoder_r_W", true), model.get("encoder_r_Wx", false))),
B_(merge(model.get("encoder_r_b", true, true), model.get("encoder_r_bx", false, true))),
U_(merge(model.get("encoder_r_U", true), model.get("encoder_r_Ux", false))),
Gamma_1_(model.get("encoder_r_gamma1", false)),
Gamma_2_(model.get("encoder_r_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::EncBackwardGRU::EncBackwardGRU(const NpzConverter& model)
: W_(model.get("encoder_r_W", true)),
B_(model.get("encoder_r_b", true, true)),
U_(model.get("encoder_r_U", true)),
Wx_(model.get("encoder_r_Wx", true)),
Bx1_(model.get("encoder_r_bx", true, true)),
Bx2_(new mblas::Tensor( Bx1_->dim(0), Bx1_->dim(1), Bx1_->dim(2), Bx1_->dim(3), true)),
Ux_(model.get("encoder_r_Ux", true)),
Gamma_1_(model.get("encoder_r_gamma1", false)),
Gamma_2_(model.get("encoder_r_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecEmbeddings::DecEmbeddings(const NpzConverter& model)
: E_(model.getFirstOfMany({std::make_pair("Wemb_dec", false),
std::make_pair("Wemb", false)}, true))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecInit::DecInit(const NpzConverter& model)
: Wi_(model.get("ff_state_W", true)),
Bi_(model.get("ff_state_b", true, true)),
Gamma_(model.get("ff_state_gamma", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecGRU1::DecGRU1(const NpzConverter& model)
: W_(model.get("decoder_W", true)),
B_(model.get("decoder_b", true, true)),
U_(model.get("decoder_U", true)),
Wx_(model.get("decoder_Wx", true)),
Bx1_(model.get("decoder_bx", true, true)),
Bx2_(new mblas::Tensor(Bx1_->dim(0), Bx1_->dim(1), Bx1_->dim(2), Bx1_->dim(3), true)),
Ux_(model.get("decoder_Ux", true)),
Gamma_1_(model.get("decoder_cell1_gamma1", false)),
Gamma_2_(model.get("decoder_cell1_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecGRU2::DecGRU2(const NpzConverter& model)
: W_(model.get("decoder_Wc", true)),
B_(model.get("decoder_b_nl", true, true)),
U_(model.get("decoder_U_nl", true)),
Wx_(model.get("decoder_Wcx", true)),
Bx2_(model.get("decoder_bx_nl", true, true)),
Bx1_(new mblas::Tensor(Bx2_->dim(0), Bx2_->dim(1), Bx2_->dim(2), Bx2_->dim(3), true)),
Ux_(model.get("decoder_Ux_nl", true)),
Gamma_1_(model.get("decoder_cell2_gamma1", false)),
Gamma_2_(model.get("decoder_cell2_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecLSTM1::DecLSTM1(const NpzConverter& model)
// matrix merging is done to be backwards-compatible with the original LSTM implementation in Amun
// we now use the same format used in Marian
// TODO: adapt to support Nematus LSTM models which use a similar format to Amun's original format
: W_(merge(model.get("decoder_W", true), model.get("decoder_Wx", false))),
B_(merge(model.get("decoder_b", true, true), model.get("decoder_bx", false, true))),
U_(merge(model.get("decoder_U", true), model.get("decoder_Ux", false))),
Gamma_1_(model.get("decoder_cell1_gamma1", false)),
Gamma_2_(model.get("decoder_cell1_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecLSTM2::DecLSTM2(const NpzConverter& model)
// matrix merging is done to be backwards-compatible with the original LSTM implementation in Amun
// we now use the same format used in Marian
// TODO: adapt to support Nematus LSTM models which use a similar format to Amun's original format
: W_(merge(model.get("decoder_Wc", true), model.get("decoder_Wcx", false))),
B_(merge(model.get("decoder_b_nl", true, true), model.get("decoder_bx_nl", false, true))),
U_(merge(model.get("decoder_U_nl", true), model.get("decoder_Ux_nl", false))),
Gamma_1_(model.get("decoder_cell2_gamma1", false)),
Gamma_2_(model.get("decoder_cell2_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecAlignment::DecAlignment(const NpzConverter& model)
: V_(model.get("decoder_U_att", true, true)),
W_(model.get("decoder_W_comb_att", true)),
B_(model.get("decoder_b_att", true, true)),
U_(model.get("decoder_Wc_att", true)),
C_(model.get("decoder_c_tt", true)), // scalar?
Gamma_1_(model.get("decoder_att_gamma1", false)),
Gamma_2_(model.get("decoder_att_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::DecSoftmax::DecSoftmax(const NpzConverter& model)
: W1_(model.get("ff_logit_lstm_W", true)),
B1_(model.get("ff_logit_lstm_b", true, true)),
W2_(model.get("ff_logit_prev_W", true)),
B2_(model.get("ff_logit_prev_b", true, true)),
W3_(model.get("ff_logit_ctx_W", true)),
B3_(model.get("ff_logit_ctx_b", true, true)),
W4_(model.getFirstOfMany({std::make_pair(std::string("ff_logit_W"), false),
std::make_pair(std::string("Wemb_dec"), true),
std::make_pair(std::string("Wemb"), true)}, true)),
B4_(model.get("ff_logit_b", true, true)),
Gamma_0_(model.get("ff_logit_l1_gamma0", false)),
Gamma_1_(model.get("ff_logit_l1_gamma1", false)),
Gamma_2_(model.get("ff_logit_l1_gamma2", false))
{}
////////////////////////////////////////////////////////////////////////////////////////////////////
Weights::Weights(const std::string& npzFile, const YAML::Node& config, unsigned device)
: Weights(NpzConverter(npzFile), config, device)
{}
Weights::Weights(const NpzConverter& model, const YAML::Node& config, unsigned device)
: encEmbeddings_(model),
decEmbeddings_(model),
decInit_(model),
decAlignment_(model),
decSoftmax_(model),
device_(device)
{
std::string encCell = config["enc-cell"] ? config["enc-cell"].as<std::string>() : "gru";
std::string encCell_r = config["enc-cell-r"] ? config["enc-cell-r"].as<std::string>() : encCell;
initEncForward(model, encCell);
initEncBackward(model, encCell_r);
std::string decCell = config["dec-cell"] ? config["dec-cell"].as<std::string>() : "gru";
std::string decCell2 = config["dec-cell-2"] ? config["dec-cell-2"].as<std::string>() : decCell;
initDec1(model, decCell);
initDec2(model, decCell2);
}
void Weights::initEncForward(const NpzConverter& model,std::string celltype){
if(celltype == "lstm"){
encForwardLSTM_ = std::shared_ptr<EncForwardLSTM>(new EncForwardLSTM(model));
} else if (celltype == "mlstm") {
encForwardMLSTM_ = std::shared_ptr<MultWeights<EncForwardLSTM>>
(new MultWeights<EncForwardLSTM>(model, "encoder"));
} else if (celltype == "gru"){
encForwardGRU_ = std::shared_ptr<EncForwardGRU>(new EncForwardGRU(model));
}
}
void Weights::initEncBackward(const NpzConverter& model,std::string celltype) {
if(celltype == "lstm"){
encBackwardLSTM_ = std::shared_ptr<EncBackwardLSTM>(new EncBackwardLSTM(model));
} else if (celltype == "mlstm") {
encBackwardMLSTM_ = std::shared_ptr<MultWeights<EncBackwardLSTM>>
(new MultWeights<EncBackwardLSTM>(model, "encoder_r"));
} else if (celltype == "gru"){
encBackwardGRU_ = std::shared_ptr<EncBackwardGRU>(new EncBackwardGRU(model));
}
}
void Weights::initDec1(const NpzConverter& model,std::string celltype){
if (celltype == "lstm"){
decLSTM1_ = std::shared_ptr<DecLSTM1>(new DecLSTM1(model));
} else if (celltype == "mlstm") {
decMLSTM1_ = std::shared_ptr<MultWeights<DecLSTM1>>(new MultWeights<DecLSTM1>(model, "decoder"));
} else if (celltype == "gru") {
decGru1_ = std::shared_ptr<DecGRU1>(new DecGRU1(model));
}
}
void Weights::initDec2(const NpzConverter& model,std::string celltype){
if (celltype == "lstm"){
decLSTM2_ = std::shared_ptr<DecLSTM2>(new DecLSTM2(model));
} else if (celltype == "mlstm") {
decMLSTM2_ = std::shared_ptr<MultWeights<DecLSTM2>>(new MultWeights<DecLSTM2>(model, "decoder_2"));
} else if (celltype == "gru") {
decGru2_ = std::shared_ptr<DecGRU2>(new DecGRU2(model));
}
}
} // namespace
}
|
