1 #include <cuda_runtime.h>
2 #include <curand.h>
3 #include <cublas_v2.h>
4
5 #include "convolutional_layer.h"
6 #include "deconvolutional_layer.h"
7 #include "gemm.h"
8 #include "blas.h"
9 #include "im2col.h"
10 #include "col2im.h"
11 #include "utils.h"
12 #include "dark_cuda.h"
13
forward_deconvolutional_layer_gpu(deconvolutional_layer layer,network_state state)14 extern "C" void forward_deconvolutional_layer_gpu(deconvolutional_layer layer, network_state state)
15 {
16 int i;
17 int out_h = deconvolutional_out_height(layer);
18 int out_w = deconvolutional_out_width(layer);
19 int size = out_h*out_w;
20
21 int m = layer.size*layer.size*layer.n;
22 int n = layer.h*layer.w;
23 int k = layer.c;
24
25 fill_ongpu(layer.outputs*layer.batch, 0, layer.output_gpu, 1);
26
27 for(i = 0; i < layer.batch; ++i){
28 float *a = layer.weights_gpu;
29 float *b = state.input + i*layer.c*layer.h*layer.w;
30 float *c = layer.col_image_gpu;
31
32 gemm_ongpu(1,0,m,n,k,1,a,m,b,n,0,c,n);
33
34 col2im_ongpu(c, layer.n, out_h, out_w, layer.size, layer.stride, 0, layer.output_gpu+i*layer.n*size);
35 }
36 add_bias_gpu(layer.output_gpu, layer.biases_gpu, layer.batch, layer.n, size);
37 activate_array(layer.output_gpu, layer.batch*layer.n*size, layer.activation);
38 }
39
backward_deconvolutional_layer_gpu(deconvolutional_layer layer,network_state state)40 extern "C" void backward_deconvolutional_layer_gpu(deconvolutional_layer layer, network_state state)
41 {
42 float alpha = 1./layer.batch;
43 int out_h = deconvolutional_out_height(layer);
44 int out_w = deconvolutional_out_width(layer);
45 int size = out_h*out_w;
46 int i;
47
48 gradient_array(layer.output_gpu, size*layer.n*layer.batch, layer.activation, layer.delta_gpu);
49 backward_bias(layer.bias_updates_gpu, layer.delta, layer.batch, layer.n, size);
50
51 if(state.delta) memset(state.delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
52
53 for(i = 0; i < layer.batch; ++i){
54 int m = layer.c;
55 int n = layer.size*layer.size*layer.n;
56 int k = layer.h*layer.w;
57
58 float *a = state.input + i*m*n;
59 float *b = layer.col_image_gpu;
60 float *c = layer.weight_updates_gpu;
61
62 im2col_ongpu(layer.delta_gpu + i*layer.n*size, layer.n, out_h, out_w,
63 layer.size, layer.stride, 0, b);
64 gemm_ongpu(0,1,m,n,k,alpha,a,k,b,k,1,c,n);
65
66 if(state.delta){
67 int m = layer.c;
68 int n = layer.h*layer.w;
69 int k = layer.size*layer.size*layer.n;
70
71 float *a = layer.weights_gpu;
72 float *b = layer.col_image_gpu;
73 float *c = state.delta + i*n*m;
74
75 gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
76 }
77 }
78 }
79
pull_deconvolutional_layer(deconvolutional_layer layer)80 extern "C" void pull_deconvolutional_layer(deconvolutional_layer layer)
81 {
82 cuda_pull_array(layer.weights_gpu, layer.weights, layer.c*layer.n*layer.size*layer.size);
83 cuda_pull_array(layer.biases_gpu, layer.biases, layer.n);
84 cuda_pull_array(layer.weight_updates_gpu, layer.weight_updates, layer.c*layer.n*layer.size*layer.size);
85 cuda_pull_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
86 }
87
push_deconvolutional_layer(deconvolutional_layer layer)88 extern "C" void push_deconvolutional_layer(deconvolutional_layer layer)
89 {
90 cuda_push_array(layer.weights_gpu, layer.weights, layer.c*layer.n*layer.size*layer.size);
91 cuda_push_array(layer.biases_gpu, layer.biases, layer.n);
92 cuda_push_array(layer.weight_updates_gpu, layer.weight_updates, layer.c*layer.n*layer.size*layer.size);
93 cuda_push_array(layer.bias_updates_gpu, layer.bias_updates, layer.n);
94 }
95
update_deconvolutional_layer_gpu(deconvolutional_layer layer,int skip,float learning_rate,float momentum,float decay)96 extern "C" void update_deconvolutional_layer_gpu(deconvolutional_layer layer, int skip, float learning_rate, float momentum, float decay)
97 {
98 int size = layer.size*layer.size*layer.c*layer.n;
99
100 axpy_ongpu(layer.n, learning_rate, layer.bias_updates_gpu, 1, layer.biases_gpu, 1);
101 scal_ongpu(layer.n, momentum, layer.bias_updates_gpu, 1);
102
103 axpy_ongpu(size, -decay, layer.weights_gpu, 1, layer.weight_updates_gpu, 1);
104 axpy_ongpu(size, learning_rate, layer.weight_updates_gpu, 1, layer.weights_gpu, 1);
105 scal_ongpu(size, momentum, layer.weight_updates_gpu, 1);
106 }
107