-
Notifications
You must be signed in to change notification settings - Fork 0
/
Network.cu
259 lines (224 loc) · 8.84 KB
/
Network.cu
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
/*
* Tanner Hoelzel
*/
#include <cmath>
#include <iostream>
#include <algorithm>
#include <vector>
#include "Network.h"
#define BATCH_SIZE 100
#define RATE 0.002
#define DO_RATE 0.4
Network::Network(float *inputs, unsigned char *labels) {
float input_w[28*28*1024];
float hidden_w[1024*10];
this->eng = new std::default_random_engine(std::random_device{}());
std::normal_distribution<float> dist;
for(unsigned int i = 0; i < 28*28*1024; i++) {
input_w[i] = dist(*(this->eng))/sqrt(28*28);
}
for(unsigned int i = 0; i < 1024*10; i++) {
hidden_w[i] = dist(*(this->eng))/sqrt(1024);
}
this->host_labels = labels;
cudaMalloc(&this->input_l, 28*28*60000*sizeof(float));
cudaMalloc(&this->input_w, 28*28*1024*sizeof(float));
cudaMalloc(&this->input_w_grad, 28*28*1024*sizeof(float));
cudaMalloc(&this->input_bias, 1024*sizeof(float));
cudaMalloc(&this->input_bias_grad, 1024*sizeof(float));
cudaMalloc(&this->hidden_l, 1024*sizeof(float));
cudaMalloc(&this->hidden_w, 1024*10*sizeof(float));
cudaMalloc(&this->hidden_w_grad, 1024*10*sizeof(float));
cudaMalloc(&this->hidden_bias, 10*sizeof(float));
cudaMalloc(&this->hidden_bias_grad, 10*sizeof(float));
cudaMalloc(&this->dropouts, 1024*sizeof(float));
cudaMalloc(&this->output_l, 10*sizeof(float));
cudaMalloc(&this->softmax_l, 10*sizeof(float));
cudaMalloc(&this->softmax_ds, 10*sizeof(float));
cudaMalloc(&this->hidden_ds, 1024*sizeof(float));
cudaMemset(this->input_bias, 0, 1024*sizeof(float));
cudaMemset(this->hidden_bias, 0, 10*sizeof(float));
cudaMemcpy(this->input_l, inputs, 60000*28*28*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(this->input_w, input_w, 28*28*1024*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(this->hidden_w, hidden_w, 1024*10*sizeof(float), cudaMemcpyHostToDevice);
}
Network::~Network() {
cudaFree(this->input_l);
cudaFree(this->input_w);
cudaFree(this->input_w_grad);
cudaFree(this->input_bias);
cudaFree(this->input_bias_grad);
cudaFree(this->hidden_l);
cudaFree(this->hidden_w);
cudaFree(this->hidden_w_grad);
cudaFree(this->hidden_bias);
cudaFree(this->hidden_bias_grad);
cudaFree(this->output_l);
cudaFree(this->softmax_l);
cudaFree(this->softmax_ds);
cudaFree(this->hidden_ds);
delete this->eng;
}
__global__ void
softmax_forward(float *input, float *output, unsigned int n) {
unsigned int i;
float max = input[0];
for(i = 1; i < n; i++) {
if(input[i] > max) max = input[i];
}
float sum = 0;
for(i = 0; i < n; i++) {
output[i] = expf(input[i] - max);
sum += output[i];
}
for(i = 0; i < n; i++) {
output[i] = output[i]/sum;
}
}
__global__ void
softmax_back(float *softmax_l, float *softmax_ds, unsigned char label) {
unsigned int id = blockIdx.x*blockDim.x + threadIdx.x;
float us = -1/softmax_l[(unsigned int)label];
if (id == (unsigned int)label) {
softmax_ds[id] = (softmax_l[(unsigned int)label]*(1 - softmax_l[id]))*us;
} else {
softmax_ds[id] = -1*softmax_l[id]*softmax_l[(unsigned int)label]*us;
}
}
__global__ void
hidden_forward(float *input, unsigned int input_size, float *weights, float *output, unsigned int output_size,
float *bias, bool relu, float *dropouts) {
int id = blockIdx.x*blockDim.x + threadIdx.x;
float dp = 0;
for (unsigned int i = 0; i < input_size; i++) {
dp += weights[id*input_size+i]*input[i];
}
dp += bias[id];
if(dropouts) {
dp *= dropouts[id];
}
output[id] = (!relu || (dp > 0))? dp : 0;
}
//TODO use multiple blocks to make this faster
__global__ void
hidden_back(float *input, unsigned int input_size, float *output, unsigned int output_size,
float *us, float *ds, float *weights, float *weights_grad, float *bias, float *bias_grad,
bool relu, float *dropout) {
unsigned int id = blockIdx.x*blockDim.x + threadIdx.x;
if (!relu || output[id] > 0) {
for (size_t j = 0; j < output_size; j++) {
if(dropout) {
if(ds) {
ds[id] += dropout[j]*us[j]*weights[j*input_size+id];
}
weights_grad[j*input_size+id] += dropout[j]*us[j]*input[id]/BATCH_SIZE;
} else {
if(ds) {
ds[id] += us[j]*weights[j*input_size+id];
}
weights_grad[j*input_size+id] += us[j]*input[id]/BATCH_SIZE;
}
if(id == 0) {
if(dropout) {
bias_grad[j] += dropout[j]*us[j]/BATCH_SIZE;
} else {
bias_grad[j] += us[j]/BATCH_SIZE;
}
}
}
}
}
__global__ void
update_weights(float *weights, float *weights_grad) {
unsigned int id = blockIdx.x*blockDim.x + threadIdx.x;
weights[id] -= weights_grad[id]*RATE;
weights_grad[id] = 0;
}
void
Network::train(unsigned int i) {
float dropouts[1024];
std::uniform_real_distribution<float> dist(0.0, 1.0);
for(unsigned int i = 0; i < 1024; i++) {
if(dist(*this->eng) < DO_RATE) {
dropouts[i] = 0;
} else {
dropouts[i] = 1/DO_RATE;
}
}
cudaMemcpy(this->dropouts, dropouts, 1024*sizeof(float), cudaMemcpyHostToDevice);
cudaMemset(this->hidden_ds, 0, 1024*sizeof(float));
hidden_forward<<<1, 1024>>>(&(this->input_l[i*28*28]), 28*28, this->input_w, this->hidden_l, 1024,
this->input_bias, true, 0);
gpu_assert(cudaDeviceSynchronize());
hidden_forward<<<1, 10>>>(this->hidden_l, 1024, this->hidden_w, this->output_l, 10,
this->hidden_bias, false, this->dropouts);
gpu_assert(cudaDeviceSynchronize());
softmax_forward<<<1, 10>>>(this->output_l, this->softmax_l, 10);
gpu_assert(cudaDeviceSynchronize());
/*
float mem[10];
cudaMemcpy(mem, this->softmax_l, 10*sizeof(float), cudaMemcpyDeviceToHost);
std::cout << (unsigned int)this->host_labels[i] << std::endl;
for(unsigned int j = 0; j < 10; j++) {
std::cout << mem[j] << " ";
}
std::cout << std::endl;
*/
softmax_back<<<1, 1>>>(this->softmax_l, this->softmax_ds, this->host_labels[i]);
gpu_assert(cudaDeviceSynchronize());
hidden_back<<<1, 1024>>>(this->hidden_l, 1024, this->output_l, 10,
this->softmax_ds, this->hidden_ds, this->hidden_w, this->hidden_w_grad,
this->hidden_bias, this->hidden_bias_grad, false, 0);
gpu_assert(cudaDeviceSynchronize());
hidden_back<<<1, 28*28>>>(&(this->input_l[i*28*28]), 28*28, this->hidden_l, 1024,
this->hidden_ds, 0, this->input_w, this->input_w_grad,
this->input_bias, this->input_bias_grad, true, this->dropouts);
gpu_assert(cudaDeviceSynchronize());
}
void
Network::train() {
std::vector<unsigned int> indices(60000);
for(unsigned int i = 0; i < 60000; i++) {
indices[i] = i;
}
std::shuffle(std::begin(indices), std::end(indices), *(this->eng));
for(unsigned int i = 0; i < (60000/BATCH_SIZE); i++) {
std::cout << "Batch " << i << std::endl;
for(unsigned int j = 0; j < BATCH_SIZE; j++) {
train(indices[i*BATCH_SIZE+j]);
}
update_weights<<<28*28, 1024>>>(this->input_w, this->input_w_grad);
gpu_assert(cudaDeviceSynchronize());
update_weights<<<1024, 10>>>(this->hidden_w, this->hidden_w_grad);
gpu_assert(cudaDeviceSynchronize());
update_weights<<<1, 1024>>>(this->input_bias, this->input_bias_grad);
gpu_assert(cudaDeviceSynchronize());
update_weights<<<1, 10>>>(this->hidden_bias, this->hidden_bias_grad);
gpu_assert(cudaDeviceSynchronize());
}
}
float
Network::test(float *tests, unsigned char *labels) {
float *d_tests;
cudaMalloc(&d_tests, 28*28*10000*sizeof(float));
cudaMemcpy(d_tests, tests, 28*28*10000*sizeof(float), cudaMemcpyHostToDevice);
unsigned int acc = 0;
for(unsigned int i = 0; i < 10000; i++) {
hidden_forward<<<1, 1024>>>(&(d_tests[i*28*28]), 28*28, this->input_w, this->hidden_l, 1024, this->input_bias, true, 0);
gpu_assert(cudaDeviceSynchronize());
hidden_forward<<<1, 10>>>(this->hidden_l, 1024, this->hidden_w, this->output_l, 10, this->hidden_bias, false, 0);
gpu_assert(cudaDeviceSynchronize());
float mem[10];
cudaMemcpy(mem, this->output_l, 10*sizeof(float), cudaMemcpyDeviceToHost);
float max = mem[0];
unsigned int max_j = 0;
for(unsigned int j = 0; j < 10; j++) {
if(mem[j] > max) {
max = mem[j];
max_j = j;
}
}
if(((unsigned int)labels[i]) == max_j) acc += 1;
}
return (float)acc/10000;
}