n_data, n_input = X.shape
n_class = np.unique(t).size
T = create_label(t, n_data, n_class)
print 'make train/test data'
n_train, n_test = 1000, 50
i = np.random.permutation(n_data)[:n_train+n_test]
i_train, i_test = np.hsplit(i, [n_train])
X_train, X_test = X[i_train, :].reshape(n_train, 1, 28, 28), X[i_test, :].reshape(n_test, 1, 28, 28)
T_train, T_test = T[i_train, :], T[i_test, :]
print 'initialize...'
linear, sigmoid, softmax, relu = act.linear(), act.sigmoid(), act.softmax(), act.relu()
conv1, conv2 = conv(20, 1, 5, 5, relu), conv(50, 20, 5, 5, relu)
pool1, pool2 = pool(2, 2, 2), pool(2, 2, 2)
neural = NN(800, 500, 10, linear, sigmoid, softmax)
error = err.cross_entropy()
cnn = CNN(conv1, pool1, conv2, pool2, neural, error)
print 'train...'
cnn.train(X_train, T_train, epsilon = 0.005, lam = 0.0001, gamma = 0.9, s_batch = 5, epochs = 50)
print 'predict...'
Y_test = cnn.predict(X_test)
accuracy = cnn.accuracy(Y_test, T_test)
print 'accuracy: {0}'.format(accuracy)
print 'save figure of loss...'
cnn.save_lossfig()
output=Dense(256,activation='relu')(output)
output=Dense(101,activation='softmax')(output)
VGG=Model(VGG.input,output)
VGG.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['accuracy'])
plot_model(VGG, to_file='./model.png', show_shapes=True, show_layer_names=True)
train_model(3, 'vgg', VGG, srgan, resize = True)
"""Test, evaluate the model"""
test_set = devide(12, 2, 2)
X = tensor2numpy('./data/', test_set, srgan)
x = [X[i] for i in X.keys()]
test = np.array(x, dtype = "float64")
y = create_onehot(X)
evaluation = cnn.evaluate(test, y)
prediction = cnn.predict(test)
"""Make predictions"""
sss = devide_submission(5)
preds = {}
for i in range(0, len(sss),5):
test_set = tensor2numpy('./data/', sss[i:i+5], model)
t = [test_set[i] for i in test_set.keys()]
test = np.array(t, dtype = 'float64')
prediction = cnn.predict(test)
top_five = np.argsort(prediction, axis=1)[:,:5]
top = mode(top_five,0)
preds[sss[i]] = top
top_five_prediction = {}
