-
Notifications
You must be signed in to change notification settings - Fork 1
/
test_layer_lenet_cls.cc
204 lines (179 loc) · 8.96 KB
/
test_layer_lenet_cls.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
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
/* test_mnist_lenet_cls.cc for LEICHT
* Copyright (C) 2017 Mo Zhou <[email protected]>
* MIT License
*/
#include <iostream>
#include "leicht.hpp"
using namespace std;
unsigned int batchsize = 64;
double lr = 1e-3; // reference lr=1e-3
int maxiter = 1000;
int iepoch = 37800/batchsize;
int itepoch = 4200/batchsize;
int overfit = 10; // (DEBUG) let it overfit on howmany batches
int testevery = 100;
string optim = "SGDM";
vector<double> validaccuhist;
vector<double> validlosshist;
Curve cv_train_loss;
Curve cv_train_acc;
Curve cv_test_loss;
Curve cv_test_acc;
int
main(void)
{
leicht_threads(2);
cout << ">> Reading MNIST training dataset" << endl;
Tensor<double> trainImages(37800, 784); trainImages.setName("trainImages");
leicht_hdf5_read("mnist.th.h5", "/train/images", 0, 0, 37800, 784, trainImages.data);
Tensor<double> trainLabels(37800, 1); trainLabels.setName("trainLabels");
leicht_hdf5_read("mnist.th.h5", "/train/labels", 0, 0, 37800, 1, trainLabels.data);
cout << ">> Reading MNIST validation dataset" << endl;
Tensor<double> valImages(4200, 784); valImages.setName("valImages");
leicht_hdf5_read("mnist.th.h5", "/val/images", 0, 0, 4200, 784, valImages.data);
Tensor<double> valLabels(4200, 1); valLabels.setName("valLabels");
leicht_hdf5_read("mnist.th.h5", "/val/labels", 0, 0, 4200, 1, valLabels.data);
cout << ">> Initialize Network" << endl;
// reference: caffe/examples/mnist/lenet
Blob<double> label (1, batchsize, "label", false);
Blob<double> X (batchsize, 784, "X", false);
Blob<double> image (batchsize, 1, 28, 28, "image", false);
Blob<double> conv1 (batchsize, 20, 24, 24); conv1.setName("conv1");
Blob<double> pool1 (batchsize, 20, 12, 12); pool1.setName("pool1");
Blob<double> conv2 (batchsize, 50, 8, 8); conv2.setName("conv2");
Blob<double> pool2 (batchsize, 50, 4, 4); pool2.setName("pool2");
Blob<double> pool2f (batchsize, 800); pool2f.setName("pool2f");
Blob<double> pool2fT (800, batchsize); pool2fT.setName("pool2fT");
Blob<double> ip1 (500, batchsize); ip1.setName("ip1");
Blob<double> ip2 (10, batchsize); ip2.setName("ip2");
Blob<double> sm1 (10, batchsize); sm1.setName("sm1");
Blob<double> loss (1); loss.setName("loss");
Blob<double> acc (1); acc.setName("acc");
Layer<double> lid1; // X->image bs,784->bs,1,28,28
Conv2dLayer<double> lconv1 (batchsize, 1, 28, 28, 20, 5); // image->conv1 bs,1,28,28->bs,20,24,24
MaxpoolLayer<double> lpool1 (batchsize, 20, 24, 24, 2, 2); // conv1->pool1 bs,20,24,24->bs,20,12,12
Conv2dLayer<double> lconv2 (batchsize, 20, 12, 12, 50, 5); // pool1->conv2 bs,20,12,12->bs,50,8,8
MaxpoolLayer<double> lpool2 (batchsize, 50, 8, 8, 2, 2); // conv2->pool2 bs,50,8,8->bs,50,4,4
Layer<double> lid2; // pool2->pool2f(lattened) bs,50,4,4->bs,800
TransposeLayer<double>lt1; // pool2f->pool2fT bs,800->800,bs
LinearLayer<double> lfc1 (500, 800); // pool2fT->ip1 800,bs->500,bs
ReluLayer<double> lrelu1; // ip1->ip1
LinearLayer<double> lfc2 (10, 500); // ip1->ip2 500,bs->10,bs
SoftmaxLayer<double> lsm1; // ip2->sm1
ClassNLLLoss<double> lloss; // sm1->loss
ClassAccuracy<double> lacc; // sm1->acc
cout << ">> Start training" << endl;
for (int iteration = 0; iteration < maxiter; iteration++) {
tic();
leicht_bar_train(iteration);
// -- get batch
X.value.copy(
//trainImages.data + (iteration%overfit)*batchsize*784, batchsize*784);
trainImages.data + (iteration%iepoch)*batchsize*784, batchsize*784);
label.value.copy(
//trainLabels.data + (iteration%overfit)*batchsize*1, batchsize*1);
trainLabels.data + (iteration%iepoch)*batchsize*1, batchsize*1);
X.value.scal_(1./255.);
// -- forward : unfold with vim: BEIGN,ENDs/; /;\r/g
lid1.forward(X, image); //X.dump(true, false); image.dump(true, false);
lconv1.forward(image, conv1); //conv1.dump(true, false);
lpool1.forward(conv1, pool1); //pool1.dump(true, false);
lconv2.forward(pool1, conv2); //conv2.dump(true, false);
lpool2.forward(conv2, pool2); //pool2.dump(true, false);
lid2.forward(pool2, pool2f); //pool2f.dump(true, false);
lt1.forward(pool2f, pool2fT); //pool2fT.dump(true, false);
//auto p2T = pool2f.value.transpose();
//pool2fT.value.copy(p2T->data, p2T->getSize());
//delete p2T;
lfc1.forward(pool2fT, ip1); //ip1.dump(true, false);
lrelu1.forward(ip1, ip1); //ip1.dump(true, false);
lfc2.forward(ip1, ip2); //ip2.dump(true, false);
lsm1.forward(ip2, sm1); //sm1.dump(true, false);
lloss.forward(sm1, loss, label); //loss.dump(true, false);
lacc.forward(sm1, loss, label); //acc.dump(true, false);
// -- zerograd
label.zeroGrad(); X.zeroGrad(); image.zeroGrad();
conv1.zeroGrad(); pool1.zeroGrad(); conv2.zeroGrad();
pool2.zeroGrad(); pool2f.zeroGrad(); pool2fT.zeroGrad();
ip1.zeroGrad(); ip2.zeroGrad(); sm1.zeroGrad();
loss.zeroGrad(); acc.zeroGrad();
lid1.zeroGrad(); lconv1.zeroGrad(); lpool1.zeroGrad();
lconv2.zeroGrad(); lpool2.zeroGrad(); lid2.zeroGrad();
lfc1.zeroGrad(); lrelu1.zeroGrad(); lfc2.zeroGrad();
lsm1.zeroGrad(); lloss.zeroGrad(); lacc.zeroGrad();
// -- backward : unfold with vim: BEIGN,ENDs/; /;\r/g
lloss.backward(sm1, loss, label); //sm1.dump();
lsm1.backward(ip2, sm1); //ip2.dump();
lfc2.backward(ip1, ip2); //ip1.dump();
lrelu1.backward(ip1, ip1); //ip1.dump();
lfc1.backward(pool2fT, ip1); //pool2fT.dump();
lt1.backward(pool2f, pool2fT); //pool2f.dump();
//auto p2fT = pool2fT.gradient.transpose();
//pool2f.gradient.copy(p2fT->data, p2fT->getSize());
// delete p2fT;
lid2.backward(pool2, pool2f); //pool2.dump();
lpool2.backward(conv2, pool2); //conv2.dump();
lconv2.backward(pool1, conv2); //pool1.dump();
lpool1.backward(conv1, pool1); //conv1.dump();
lconv1.backward(image, conv1); //image.dump();
// regularize
lconv1.regularization(); lconv2.regularization();
lfc1.regularization(); lfc2.regularization();
// -- report
lloss.report(); lacc.report(true);
label.dump(true, false);
lconv1.dumpstat(); lconv2.dumpstat();
lfc1.dumpstat(); lfc2.dumpstat();
//pool1.dump(true, false);
cv_train_loss.append(iteration, lloss.lossval);
cv_train_acc.append(iteration, lacc.accuracy);
// -- update
lconv1.update(lr, optim); lconv2.update(lr, optim);
lfc1.update(lr, optim); lfc2.update(lr, optim);
toc();
// -- validation
if (testevery!=0 && iteration%testevery==0) {
leicht_bar_val(iteration);
Tensor<double> cvloss (itepoch);
Tensor<double> cvacc (itepoch);
for (int t = 0; t < itepoch; t++) {
// -- get batch
X.value.copy(valImages.data + t*batchsize*784, batchsize*784);
label.value.copy(valLabels.data + t*batchsize*1, batchsize*1);
X.value.scal_(1./255.);
// -- forward : unfold with vim: BEIGN,ENDs/; /;\r/g
lid1.forward(X, image); //X.dump(true, false); image.dump(true, false);
lconv1.forward(image, conv1); //conv1.dump(true, false);
lpool1.forward(conv1, pool1); //pool1.dump(true, false);
lconv2.forward(pool1, conv2); //conv2.dump(true, false);
lpool2.forward(conv2, pool2); //pool2.dump(true, false);
lid2.forward(pool2, pool2f); //pool2f.dump(true, false);
lt1.forward(pool2f, pool2fT); //pool2fT.dump(true, false);
//auto p2T = pool2f.value.transpose();
//pool2fT.value.copy(p2T->data, p2T->getSize());
//delete p2T;
lfc1.forward(pool2fT, ip1); //ip1.dump(true, false);
lrelu1.forward(ip1, ip1); //ip1.dump(true, false);
lfc2.forward(ip1, ip2); //ip2.dump(true, false);
lsm1.forward(ip2, sm1); //sm1.dump(true, false);
lloss.forward(sm1, loss, label); //loss.dump(true, false);
lacc.forward(sm1, loss, label); //acc.dump(true, false);
// -- report
//lloss.report(); lacc.report();
cout << "."; cout.flush();
cvloss.data[t] = lloss.lossval;
cvacc.data[t] = lacc.accuracy;
}
cout << endl;
cout << "Test Loss" << cvloss.sum() / cvloss.getSize() << endl;
cv_test_loss.append(iteration, cvloss.sum() / cvloss.getSize());
cout << "Test Accu" << cvacc.sum() / cvacc.getSize() << endl;
cv_test_acc.append(iteration, cvacc.sum() / cvacc.getSize());
}
}
cv_train_loss.draw("lenet-train-loss.svg");
cv_train_acc.draw("lenet-train-acc.svg");
cv_test_loss.draw("lenet-test-loss.svg");
cv_test_acc.draw("lenet-test-acc.svg");
return 0;
}