mnist_nn_keras.py

from keras.datasets import mnist
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras.utils import np_utils
import numpy as np

l1_nodes = 200
l2_nodes = 100
final_layer_nodes = 10

(X_train, y_train), (X_test, y_test) = mnist.load_data()

X_train = X_train.reshape(60000, 784).astype("float32")
X_test = X_test.reshape(10000, 784).astype("float32")
X_train /= 255
X_test /= 255

Y_train = np_utils.to_categorical(y_train, nb_classes)
Y_test = np_utils.to_categorical(y_test, nb_classes)

# model
model = Sequential()
model.add(Dense(units=l1_nodes, input_shape=(784, ), activation="relu"))
model.add(Dense(units=l2_nodes, input_shape=(l1_nodes, ), activation="relu", ))
model.add(Dense(units=final_layer_nodes, input_shape=(l2_nodes, )))
model.add(Activation("softmax"))

model.compile(loss='categorical_crossentropy',
              optimizer='sgd',
              metrics=["accuracy"])

model.fit(X_train, Y_train,
          batch_size=100,
          epochs=4,
          verbose=1,
          validation_data=(X_test, Y_test))