Keras neural network outputs same result for every input

Question

I tried to implement a feedforward neural network.

This is the structure: Input layer: 8 neurons, Hidden layer: 8 neurons and Output layer: 8 neurons.

The input data are vectors of 8 bits (1 bit for each neuron of the input layer). The outputs of the neural network are also vectors of 8 bits. So in total the dataset has 256 examples.

Example: if given x = [0.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0]

the output must be y = [1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 1.0]

This is the implementation:

from keras.models import Sequential
from keras.layers import Dense
import numpy as np
import random
from math import ceil

#Dimension of layers
dim = 8

#Generate dataset
X = []
for i in range(0,2**dim):
    n = [float(x) for x in bin(i)[2:]]
    X.append([0.]*(dim-len(n))+n)
y = X[:]
random.shuffle(y)
X = np.array(X)
y = np.array(y)

# create model
model = Sequential()
model.add(Dense(dim, input_dim=dim, init='normal', activation='sigmoid'))
model.add(Dense(dim, init='normal', activation='sigmoid'))
model.add(Dense(dim, init='normal', activation='sigmoid'))

# Compile model
model.compile(loss='mse', optimizer='SGD', metrics=['accuracy'])
# Fit the model
model.fit(X, y, nb_epoch=1000, batch_size=50, verbose=0)
# evaluate the model
scores = model.evaluate(X, y)
print("%s: %.2f%%" % (model.metrics_names[1], scores[1]*100))
output = model.predict(X)

#Make the output binary
for i in range(0, output[:,0].size):
    for j in range(0, output[0].size):
        if output[i][j] > 0.5 or output[i][j] == 0.5:
            output[i][j] = 1
        else:
            output[i][j] = 0
print(output)

This is what I get in output:

acc: 50.39%
[[ 1.  0.  0. ...,  0.  1.  1.]
[ 1.  0.  0. ...,  0.  1.  1.]
[ 1.  0.  0. ...,  0.  1.  1.]
..., 
[ 1.  0.  0. ...,  0.  1.  1.]
[ 1.  0.  0. ...,  0.  1.  1.]
[ 1.  0.  0. ...,  0.  1.  1.]]

It seems that all outputs have the same value. So I don´t know what's wrong about the configuration. I tried this Cannot train a neural network in keras - stackoverflow which suggests to remove the activation function at the output layer but when I run this I get all output vectors with this value:

[ 0. 1. 1. ..., 1. 1. 1.]

Any insights on how to make it work?

Solution 1

I had the very same problem.

I would suggest you to reduce the learning rate for SGD. In my case I had used Adam Optimizer with lr=0.001, but changing to 0.0001 solved the problem.

Default parameters for SGD are:

keras.optimizers.SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=False)

Solution 2

The output is relatively similar to multi-label classification so I would recommend:

1. Change loss function to binary_crossentropy
2. Retain the last activation layer as sigmoid and change the others - relu can be a good choice.
3. Add validation to your "fit" call and increase verbosity - This will allow you to understand how your network changes through the epochs and especially when it over/under fits
4. Add depth to the network until you overfit
5. Add regularization to your network until you don't overfit
6. repeat 4+5

Solution 3

If you tried all the above and it does not work it means that you try to fit noise, there is no connection/correlation/relevance between your inputs and outputs.

Comments

[xtluo]

This solved my problem, which is my CNN produces the same probability distribution for all inputs...confused!