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
|
#pragma once
#include "graph/expression_graph.h"
#include "graph/node_initializers.h"
namespace marian {
Expr debug(Expr a, const std::string& message = "");
typedef Expr(ActivationFunction)(Expr);
Expr plus(const std::vector<Expr>&);
// TODO: should be logistic(), not sigmoid()
Expr sigmoid(Expr a);
Expr sigmoid(const std::vector<Expr>&);
Expr swish(Expr a);
Expr swish(const std::vector<Expr>&);
Expr tanh(const std::vector<Expr>&);
template <typename... Args>
Expr tanh(Args... args) {
std::vector<Expr> nodes{args...};
return tanh(nodes);
}
Expr relu(Expr a);
Expr relu(const std::vector<Expr>&);
Expr leakyrelu(Expr a);
Expr leakyrelu(const std::vector<Expr>&);
Expr prelu(Expr a, float alpha = 0.01);
Expr prelu(const std::vector<Expr>&, float alpha = 0.01);
Expr log(Expr a);
Expr exp(Expr a);
Expr clip(Expr a, float c);
Expr operator-(Expr a);
/*********************************************************/
Expr operator+(Expr a, Expr b);
Expr operator+(float a, Expr b);
Expr operator+(Expr a, float b);
Expr operator-(Expr a, Expr b);
Expr operator-(float a, Expr b);
Expr operator-(Expr a, float b);
Expr operator*(Expr a, Expr b);
Expr operator*(float a, Expr b);
Expr operator*(Expr a, float b);
Expr operator/(Expr a, Expr b);
Expr operator/(float a, Expr b);
Expr operator/(Expr a, float b);
// Expr pow(Expr a, Expr b);
// Expr pow(float a, Expr b);
// Expr pow(Expr a, float b);
Expr logaddexp(Expr a, Expr b);
// Note: Following numpy, minimum() is element-wise, while min() is along an axis in both Numpy and PyTorch.
Expr maximum(Expr a, Expr b);
Expr minimum(Expr a, Expr b);
// Note: We cannot overload the relational operators, as they also mean something for Expr itself.
// Note: These names follow PyTorch convention.
Expr lt(Expr a, Expr b);
Expr eq(Expr a, Expr b);
Expr gt(Expr a, Expr b);
Expr ge(Expr a, Expr b);
Expr ne(Expr a, Expr b);
Expr le(Expr a, Expr b);
Expr lt(float a, Expr b);
Expr eq(float a, Expr b);
Expr gt(float a, Expr b);
Expr ge(float a, Expr b);
Expr ne(float a, Expr b);
Expr le(float a, Expr b);
Expr lt(Expr a, float b);
Expr eq(Expr a, float b);
Expr gt(Expr a, float b);
Expr ge(Expr a, float b);
Expr ne(Expr a, float b);
Expr le(Expr a, float b);
Expr dot(Expr a,
Expr b,
bool transA = false,
bool transB = false,
float scalar = 1.f);
Expr bdot(Expr a,
Expr b,
bool transA = false,
bool transB = false,
float scalar = 1.f);
Expr affine(Expr a,
Expr b,
Expr c,
bool transA = false,
bool transB = false,
float scalar = 1.f);
Expr csr_dot(const Shape& A_shape, Expr Avalues, Expr Aindices, Expr Aoffsets, Expr B, bool transA = false);
Expr dot_csr(Expr A, const Shape& B_shape, Expr B_values, Expr B_indices, Expr B_offsets, bool transB = false);
Expr transpose(Expr a);
Expr transpose(Expr a, const std::vector<int>& axes);
Expr swapAxes(Expr x, int axis1, int axis2);
Expr concatenate(const std::vector<Expr>& concats, int ax = 0);
Expr repeat(Expr a, size_t repeats, int ax = 0);
Expr reshape(Expr a, Shape shape);
Expr atleast_1d(Expr a);
Expr atleast_2d(Expr a);
Expr atleast_3d(Expr a);
Expr atleast_4d(Expr a);
Expr atleast_nd(Expr a, size_t dims);
// create a constant of shape a->shape() and initialize with init
Expr constant_like(Expr a, const NodeInitializer& init);
Expr flatten(Expr a);
Expr flatten_2d(Expr a);
Expr stopGradient(Expr a);
Expr gather(Expr a, int axis, Expr indices);
// Warning: Don't try to pass a scalar literal 0 as indices; it will compile but pass nullptr...
Expr index_select(Expr a, int axis, Expr indices);
// convenience wrappers for index_select()
Expr index_select(Expr a, int axis, const std::vector<IndexType>& indices);
static inline Expr rows(Expr a, Expr indices) {
return index_select(a, 0, indices);
}
static inline Expr rows(Expr a, const std::vector<IndexType>& indexVector) {
return index_select(a, 0, indexVector);
}
static inline Expr cols(Expr a, Expr indices) {
return index_select(a, -1, indices);
}
static inline Expr cols(Expr a, const std::vector<IndexType>& indexVector) {
return index_select(a, -1, indexVector);
}
Expr slice(Expr a, int axis, Slice slice);
// convenience wrappers for slice()
static inline Expr slice(Expr a, int axis, int index) { // single index @NOTE: This was formerlly called step()
return slice(a, axis, Slice(index));
}
static inline Expr narrow(Expr a, int axis, size_t start, size_t length) { // PyTorch name
return slice(a, axis, Slice((int)start, (int)(start + length)));
}
/*********************************************************/
Expr sum(Expr a, int ax = 0);
Expr mean(Expr a, int ax = 0);
Expr std(Expr a, int ax);
Expr var(Expr a, int ax);
Expr max(Expr a, int ax);
Expr min(Expr a, int ax);
Expr prod(Expr a, int ax);
Expr logsumexp(Expr a, int ax);
Expr softmax(Expr x, int axis = -1);
// @TODO: maybe get rid of this entirely to not obfuscate, what's going on inside.
// @TODO: switch to log-masking everywhere?
Expr softmax(Expr a, Expr zeroOneMask, int axis = -1);
Expr logsoftmax(Expr a);
Expr cross_entropy(Expr a, Expr b);
Expr scalar_product(Expr a, Expr b, int ax = 0);
Expr weighted_average(Expr in, Expr weights, int ax = 0);
Expr sqrt(Expr a, float eps = 0.f);
Expr square(Expr a);
Expr layerNorm(Expr x, Expr gamma, Expr beta = nullptr, float eps = 1e-9);
Expr highway(Expr y, Expr x, Expr t);
Expr highway(const std::string prefix, Expr x);
static inline Expr dropout(Expr x, Expr mask) {
return x * mask;
}
static inline Expr dropout(Expr x, float dropProb, Shape shape) {
if(dropProb == 0)
return x;
auto graph = x->graph();
auto mask = graph->dropout(dropProb, shape);
return dropout(x, mask);
}
static inline Expr dropout(Expr x, float dropProb) {
if(dropProb == 0)
return x;
return dropout(x, dropProb, x->shape());
}
Expr shift(Expr, Shape, float padValue = 0);
Expr convert2cudnnFormat(Expr x);
Expr convertFromcudnnFormat(Expr x);
Expr avg_pooling(Expr x,
int height,
int width,
int padHeight = 0,
int padWidth = 0,
int strideHeight = 1,
int strideWidth = 1);
Expr max_pooling(Expr x,
int height,
int width,
int padHeight = 0,
int padWidth = 0,
int strideHeight = 1,
int strideWidth = 1);
Expr pooling_with_masking(Expr x, Expr mask, int width, bool isEven = false);
} // namespace marian
|
