ReLU not learning to handle negative inputs Keras

问题:

I want my neural network to convert a negative value into a positive value. Theoretically this can be done using a ReLU function and 1 node which learns the input weight to be -1 (so a negative input is multiplied by -1 = positive input.

It just keeps on outputting 0. Code below. I used -1 as input values to see if it could learn on at least a single input.

I tried adding more layers but it doesn't help see edit, IT DID help if I add more

train_input = np.asarray([[-1]]*10000) # Input arr of -1s
train_output = np.asarray(map(lambda x: [abs(x[0])] , train_input))

# Define the model
model = Sequential()
model.add(Dense(1, input_dim=1, kernel_initializer='normal', activation='linear'))
model.add(LeakyReLU(alpha=.001))
model.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])

# Train and evaluate
model.fit(train_input, train_output, epochs=10, batch_size=10, verbose=0)
test_model_output = model.predict(test_input)
print str(test_input[0][0]) + " " + str(test_output[0][0]) + " " +  str(test_model_output[0][0])

The output I get is below (1st value is input, 2nd is expected output, 3rd is model output )

-1 1 0.0

EDIT I tried using the random uniform initialiser so it would initialise negative weights and it works. I get why this should make it easier for the network to learn. But I don't get why it's necessary.

from keras.initializers import RandomUniform
model.add(Dense(1, input_dim=1, kernel_initializer=RandomUniform(minval=-0.05, maxval=0.05, seed=None), activation='linear'))

EDIT 2 Someone mentioned I didn't have enough time to train the data. At first I thought making it 10x more data and batches to be 10x smaller (more iterations) would work. It didn't BUT if I added 10x more epochs (100 total) it did work. So it just takes a long time to convert positive initialised weights to negative

回答1:

I would initialise that first weight to be negative using keras.initializers.Constant(value=-1) https://keras.io/initializers/#constant

May help to get that first neuron firing.

回答2:

I'm not familiar with the libraries you're using, but it seems like you have to speicify the input_shape in the Dense line.

where None indicates that any positive integer may be expected

So if you input nothing as input_shape, then it assumes your input will be any positive number.

So adding input_shape=(-1,1) might fix your problem!

https://keras.io/getting-started/sequential-model-guide/

回答3:

I think the outcome of the model is correct.

Because​, Rectified Linear Unit works as follows

f(x) = max(0,x) [x - > input to the neuron]

In your example, input value is -1. i.e (x = -1)

f(x) = max(0,-1) => Answer for this equation is Zero ['0']

This may be the reason behind your models' result

(or)

Possibly, error while performing multiplication may occurred.

回答4:

The problem was that I was not giving it enough time to train. While it's a very basic function the initialised weights have to go form negative to positive.

Increase the amount of training done (more epochs, smaller batches, more training data) eventually caused the gradient to flip from positive to negative.