def main(backend):
(X_train, y_train), (X_test, y_test) = get_mnist(backend)
# set the backend
nn = net.set_backend(backend)
model = nn.Sequential()
model.add(nn.Dense(512, input_shape=(784,)))
model.add(nn.Activation('relu'))
model.add(nn.BatchNormalization())
model.add(nn.Dropout(p=0.2))
model.add(nn.Dense(512))
model.add(nn.Activation('relu'))
model.add(nn.BatchNormalization())
model.add(nn.Dropout(p=0.2))
model.add(nn.Dense(10))
model.add(nn.Activation('softmax'))
rms = nn.Adam()
model.compile(loss='categorical_crossentropy', optimizer=rms)
model.fit(X_train, y_train, nb_epoch=20)#, optimizer=rms)
proba = model.predict(X_test)
pred = np.argmax(proba, axis=1)
if backend == 'keras':
print 'Test error: {:.05f}'.format(np.mean(pred == np.argmax(y_test, axis=1)))
else:
print 'Test error: {:.05f}'.format(np.mean(pred == y_test))
return model
def main(backend):
print('Loading data...')
X, labels = load_data('train.csv', train=True)
X, scaler = preprocess_data(X)
if backend == 'keras':
y, encoder = preprocess_labels(labels)
nb_classes = y.shape[1]
print(nb_classes, 'classes')
else:
y, encoder = preprocess_labels(labels, categorical=False)
nb_classes = len(np.unique(y))
print(nb_classes, 'classes')
dims = X.shape[1]
print(dims, 'dims')
nn = net.set_backend(backend)
model = nn.Sequential()
model.add(nn.Dense(512, input_shape=(dims,)))
model.add(nn.PReLU())
model.add(nn.BatchNormalization())
model.add(nn.Dropout(p=0.5))
model.add(nn.Dense(512))
model.add(nn.PReLU())
model.add(nn.BatchNormalization())
model.add(nn.Dropout(p=0.5))
model.add(nn.Dense(512))
model.add(nn.PReLU())
model.add(nn.BatchNormalization())
model.add(nn.Dropout(p=0.5))
model.add(nn.Dense(nb_classes))
model.add(nn.Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam')
model.fit(X, y, nb_epoch=20, batch_size=128, validation_split=0.15)
return model
