Use keras to build a simple ANN but get 0 accuarcy

Question

I'm learning keras, and my task is simple. Use 3 vars from data to predict another var value, it works on R/Python ANN models with ok accuracy. But when I tried to use keras to build a neural network, it failed to work with 0 accuracy. The code is,

from keras.models import Sequential
from keras.layers import Dense, Activation
import pandas as pd
from sklearn.cross_validation import train_test_split

data = pd.read_csv(datapath)
data = data.dropna()

x = data.values[:, [2, 9, 10]]
y = data.values[:, 8]
train_X, test_X, train_y, test_y = train_test_split(x, y, train_size=0.5, random_state=0)

model = Sequential()
model.add(Dense(16, input_shape=(3,)))
model.add(Activation('sigmoid'))
model.add(Dense(1))
model.add(Activation('linear'))

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

model.fit(train_X, train_y, nb_epoch=100, batch_size=32, verbose=0)
loss, accuracy = model.evaluate(test_X, test_y, verbose=0)
print("Accuracy = {:.2f}".format(accuracy))

So, my question is, is my code right and How to make it work?

Answer 1

If you try to predict a value which is real number, you might not be interested in the accuracy which is dedicated for classification. I would suggest to measure the training progress with the loss you are using : mean square error.

If you want to monitor training progress, you can pass a keyword validation_data=[X_val, y_val] to the fit method.

Otherwise, to measure the prediction error on the testing set just use the scikit-learn function http://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_squared_error.html

Update One difference with other packages might be the initialization of the weights. Playing with the initialization parameter of Dense layers or the scale of initialization might help you.