Keras输入/输出

Question

I am struggling with a challenge in Tensorflow / keras, would be great if someone could help me.

I have build a neural net in Keras with input_dim=3, then 10 Neurons and Output 1.

The input is a 3d-vector with floats, the output should be a simple float value.

My problem is, that I dont know how the floats should be formatted (>1, from 0 to 1?, etc...) and which loss function could work out for this task (nothing binary i guess). I want the neural net to compute out of the 3d vector a simple float value. But it never works out because my outputs are always the same.

If I have forgotten something please let me know, if you have some ideas to it, it would be great!

Greetings

Edit: Im aware of the fact that I need an introduction into the whole topic of machine learning, which I am doing right now. In the mean time I would like to know how to use keras to verifiy/practically use machine learning. I am sorry for asking 'stupid' questions but I hope that maybe someone could help me.

Input: I think the input might be 'wrong' formatted, its not normalized etc., but I transformed the values i get to an interval mentioned below.

This is my simple model:

model = Sequential()
model.add(Dense(10, input_dim=3, init='normal', activation='sigmoid'))
model.add(Dense(1, init='normal', activation='sigmoid'))
model.compile(loss='mse', optimizer='sgd', metrics=['accuracy'])
model.fit(X_Train, Y_Train, nb_epoch=100, batch_size=32, verbose=1)

X_Train and Y_Train are values extracted from a .csv file. For example my input values are [a,b,c,d], where 0 < a,b,c < 1 and -1 < d < 1 (d is output).

Output:

Epoch 500/500

32/32 [==============================] - 0s - loss: 0.0813 - acc: 0.0000e+00

Example (random generated values), all output is nearly the same around 0.43:

[ 0.97650245 0.30383579 0.74829968] [[ 0.43473071]]

[ 0.94985165 0.75347051 0.72609185] [[ 0.43473399]]

[ 0.18072594 0.18540003 0.20763266] [[ 0.43947196]]

Answer 1

Firstly, there is no need for normalizing (or formatting) the input samples.

Secondly, for the problem of zero accuracy, it's because you used the "accuracy" as metrics which is used for classification model. In your case, you should use something like "mse" or "mae" (in Keras) as metrics in your compile method, eg,

model.compile(loss='mse', optimizer='sgd', metrics=['mae'])

Answer 2

I'm answering my own question:

The problem here is the optimizer! The training data and every other setting is not that important. You have to try other optimizers, to vary the results. Its possible to close this question now. Thank you for your help!

Answer 3

Change your output layer to following model.add(Dense(1))

See this regression guide that talks about one output. https://www.tensorflow.org/tutorials/keras/basic_regression