-
Notifications
You must be signed in to change notification settings - Fork 11
/
demo.cc
98 lines (94 loc) · 3.09 KB
/
demo.cc
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
/*
* CNN demo for MNIST dataset
* Author: Kai Han ([email protected])
* Details in https://github.com/iamhankai/mini-dnn-cpp
* Copyright 2018 Kai Han
*/
#include <Eigen/Dense>
#include <algorithm>
#include <iostream>
#include "src/layer.h"
#include "src/layer/conv.h"
#include "src/layer/fully_connected.h"
#include "src/layer/ave_pooling.h"
#include "src/layer/max_pooling.h"
#include "src/layer/relu.h"
#include "src/layer/sigmoid.h"
#include "src/layer/softmax.h"
#include "src/loss.h"
#include "src/loss/mse_loss.h"
#include "src/loss/cross_entropy_loss.h"
#include "src/mnist.h"
#include "src/network.h"
#include "src/optimizer.h"
#include "src/optimizer/sgd.h"
int main() {
// data
MNIST dataset("../data/mnist/");
dataset.read();
int n_train = dataset.train_data.cols();
int dim_in = dataset.train_data.rows();
std::cout << "mnist train number: " << n_train << std::endl;
std::cout << "mnist test number: " << dataset.test_labels.cols() << std::endl;
// dnn
Network dnn;
Layer* conv1 = new Conv(1, 28, 28, 4, 5, 5, 2, 2, 2);
Layer* pool1 = new MaxPooling(4, 14, 14, 2, 2, 2);
Layer* conv2 = new Conv(4, 7, 7, 16, 5, 5, 1, 2, 2);
Layer* pool2 = new MaxPooling(16, 7, 7, 2, 2, 2);
Layer* fc3 = new FullyConnected(pool2->output_dim(), 32);
Layer* fc4 = new FullyConnected(32, 10);
Layer* relu1 = new ReLU;
Layer* relu2 = new ReLU;
Layer* relu3 = new ReLU;
Layer* softmax = new Softmax;
dnn.add_layer(conv1);
dnn.add_layer(relu1);
dnn.add_layer(pool1);
dnn.add_layer(conv2);
dnn.add_layer(relu2);
dnn.add_layer(pool2);
dnn.add_layer(fc3);
dnn.add_layer(relu3);
dnn.add_layer(fc4);
dnn.add_layer(softmax);
// loss
Loss* loss = new CrossEntropy;
dnn.add_loss(loss);
// train & test
SGD opt(0.001, 5e-4, 0.9, true);
// SGD opt(0.001);
const int n_epoch = 5;
const int batch_size = 128;
for (int epoch = 0; epoch < n_epoch; epoch ++) {
shuffle_data(dataset.train_data, dataset.train_labels);
for (int start_idx = 0; start_idx < n_train; start_idx += batch_size) {
int ith_batch = start_idx / batch_size;
Matrix x_batch = dataset.train_data.block(0, start_idx, dim_in,
std::min(batch_size, n_train - start_idx));
Matrix label_batch = dataset.train_labels.block(0, start_idx, 1,
std::min(batch_size, n_train - start_idx));
Matrix target_batch = one_hot_encode(label_batch, 10);
if (false && ith_batch % 10 == 1) {
std::cout << ith_batch << "-th grad: " << std::endl;
dnn.check_gradient(x_batch, target_batch, 10);
}
dnn.forward(x_batch);
dnn.backward(x_batch, target_batch);
// display
if (ith_batch % 50 == 0) {
std::cout << ith_batch << "-th batch, loss: " << dnn.get_loss()
<< std::endl;
}
// optimize
dnn.update(opt);
}
// test
dnn.forward(dataset.test_data);
float acc = compute_accuracy(dnn.output(), dataset.test_labels);
std::cout << std::endl;
std::cout << epoch + 1 << "-th epoch, test acc: " << acc << std::endl;
std::cout << std::endl;
}
return 0;
}