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
| #include "connected_layer.h"
| #include "utils.h"
| #include "cuda.h"
| #include "blas.h"
| #include "gemm.h"
|
| #include <math.h>
| #include <stdio.h>
| #include <stdlib.h>
| #include <string.h>
|
| connected_layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation)
| {
| int i;
| connected_layer l = {0};
| l.type = CONNECTED;
|
| l.inputs = inputs;
| l.outputs = outputs;
| l.batch=batch;
|
| l.output = calloc(batch*outputs, sizeof(float*));
| l.delta = calloc(batch*outputs, sizeof(float*));
|
| l.weight_updates = calloc(inputs*outputs, sizeof(float));
| l.bias_updates = calloc(outputs, sizeof(float));
|
| l.weights = calloc(inputs*outputs, sizeof(float));
| l.biases = calloc(outputs, sizeof(float));
|
|
| float scale = 1./sqrt(inputs);
| for(i = 0; i < inputs*outputs; ++i){
| l.weights[i] = 2*scale*rand_uniform() - scale;
| }
|
| for(i = 0; i < outputs; ++i){
| l.biases[i] = scale;
| }
|
| #ifdef GPU
| l.weights_gpu = cuda_make_array(l.weights, inputs*outputs);
| l.biases_gpu = cuda_make_array(l.biases, outputs);
|
| l.weight_updates_gpu = cuda_make_array(l.weight_updates, inputs*outputs);
| l.bias_updates_gpu = cuda_make_array(l.bias_updates, outputs);
|
| l.output_gpu = cuda_make_array(l.output, outputs*batch);
| l.delta_gpu = cuda_make_array(l.delta, outputs*batch);
| #endif
| l.activation = activation;
| fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
| return l;
| }
|
| void update_connected_layer(connected_layer l, int batch, float learning_rate, float momentum, float decay)
| {
| axpy_cpu(l.outputs, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
| scal_cpu(l.outputs, momentum, l.bias_updates, 1);
|
| axpy_cpu(l.inputs*l.outputs, -decay*batch, l.weights, 1, l.weight_updates, 1);
| axpy_cpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
| scal_cpu(l.inputs*l.outputs, momentum, l.weight_updates, 1);
| }
|
| void forward_connected_layer(connected_layer l, network_state state)
| {
| int i;
| for(i = 0; i < l.batch; ++i){
| copy_cpu(l.outputs, l.biases, 1, l.output + i*l.outputs, 1);
| }
| int m = l.batch;
| int k = l.inputs;
| int n = l.outputs;
| float *a = state.input;
| float *b = l.weights;
| float *c = l.output;
| gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
| activate_array(l.output, l.outputs*l.batch, l.activation);
| }
|
| void backward_connected_layer(connected_layer l, network_state state)
| {
| int i;
| gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);
| for(i = 0; i < l.batch; ++i){
| axpy_cpu(l.outputs, 1, l.delta + i*l.outputs, 1, l.bias_updates, 1);
| }
| int m = l.inputs;
| int k = l.batch;
| int n = l.outputs;
| float *a = state.input;
| float *b = l.delta;
| float *c = l.weight_updates;
| gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);
|
| m = l.batch;
| k = l.outputs;
| n = l.inputs;
|
| a = l.delta;
| b = l.weights;
| c = state.delta;
|
| if(c) gemm(0,1,m,n,k,1,a,k,b,k,0,c,n);
| }
|
| #ifdef GPU
|
| void pull_connected_layer(connected_layer l)
| {
| cuda_pull_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
| cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
| cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
| cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
| }
|
| void push_connected_layer(connected_layer l)
| {
| cuda_push_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
| cuda_push_array(l.biases_gpu, l.biases, l.outputs);
| cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
| cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
| }
|
| void update_connected_layer_gpu(connected_layer l, int batch, float learning_rate, float momentum, float decay)
| {
| axpy_ongpu(l.outputs, learning_rate/batch, l.bias_updates_gpu, 1, l.biases_gpu, 1);
| scal_ongpu(l.outputs, momentum, l.bias_updates_gpu, 1);
|
| axpy_ongpu(l.inputs*l.outputs, -decay*batch, l.weights_gpu, 1, l.weight_updates_gpu, 1);
| axpy_ongpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates_gpu, 1, l.weights_gpu, 1);
| scal_ongpu(l.inputs*l.outputs, momentum, l.weight_updates_gpu, 1);
| }
|
| void forward_connected_layer_gpu(connected_layer l, network_state state)
| {
| int i;
| for(i = 0; i < l.batch; ++i){
| copy_ongpu_offset(l.outputs, l.biases_gpu, 0, 1, l.output_gpu, i*l.outputs, 1);
| }
| int m = l.batch;
| int k = l.inputs;
| int n = l.outputs;
| float * a = state.input;
| float * b = l.weights_gpu;
| float * c = l.output_gpu;
| gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
| activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
| }
|
| void backward_connected_layer_gpu(connected_layer l, network_state state)
| {
| int i;
| gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
| for(i = 0; i < l.batch; ++i){
| axpy_ongpu_offset(l.outputs, 1, l.delta_gpu, i*l.outputs, 1, l.bias_updates_gpu, 0, 1);
| }
| int m = l.inputs;
| int k = l.batch;
| int n = l.outputs;
| float * a = state.input;
| float * b = l.delta_gpu;
| float * c = l.weight_updates_gpu;
| gemm_ongpu(1,0,m,n,k,1,a,m,b,n,1,c,n);
|
| m = l.batch;
| k = l.outputs;
| n = l.inputs;
|
| a = l.delta_gpu;
| b = l.weights_gpu;
| c = state.delta;
|
| if(c) gemm_ongpu(0,1,m,n,k,1,a,k,b,k,0,c,n);
| }
| #endif
|
|
