# mnist_train.py
# 손글씨 분류 신경망 학습
from ShinNet_relu import Inet1, Inet2 # Inet2(lr=0.5) => 97%
#from ShinNet_sigmoid import Inet1, Inet2 # Inet2(lr=0.1)=> 92%
from mnist_ready import load_mnist # load dataset from mnist_ready.py
mnist = load_mnist(normalize=True, one_hot_label=False)
# group 나누기: X_train, t_train, X_test, t_test
X_train, t_train = mnist['train_img'], mnist['train_label']
X_test, t_test = mnist['test_img'], mnist['test_label']
img_size = X_train.shape[1]
N_clss = 10
batch_size = 100
N_epochs = 20
''' One layer NN '''
layer1 = 0
if layer1 == 1:
Net = Inet1(i_size=img_size, o_size=N_clss)
learning_rate = .02 # Inet1
print("---- Classification for Mnist using One-layer ANN training. ---- ")
for epoch in range(1, N_epochs+1):
Net.train(X_train, t_train, learning_rate, batch_size)
acc1 = Net.accuracy(X_train, t_train)
acc2 = Net.accuracy(X_test, t_test)
print('After {}-th epoch, Probability for Train({}), Test({})' \
.format(epoch,acc1, acc2) )
if (acc1+acc2)/2 > 0.97: break
print('')
layer2 = 1
if layer2 == 1:
''' Two layer NN '''
Net = Inet2(i_size=img_size, h_size=100, o_size=N_clss)
learning_rate = .5 # relu
# learning_rate = 0.1 # sigmoid
print("---- Classification for Mnist using Two-layer ANN training. ---- ")
for epoch in range(1, N_epochs+1):
Net.train(X_train, t_train, learning_rate, batch_size)
acc1 = Net.accuracy(X_train, t_train)
acc2 = Net.accuracy(X_test, t_test)
print('After {}-th epoch, Probability for Train({}), Test({})' \
.format(epoch,acc1, acc2) )
if (acc1+acc2)/2 > 0.97: break
|
  |
| LAST UPDATE: 2019.07.10 - 15:37 |
|
메일보내기
이름검색
신경망 파일: 1-layer(Inet1), 2-layer(Inet2)
손글씨 준비
