CODEWITHSUNDEEP
pythonkerasneural-network
Arkleseisure
Arkleseisure

Reputation: 438

Output in Keras has too many dimensions

I'm trying to make a neural network in Keras, but in doing so have come across a problem: the output size of the network isn't the same as the output I'd like, which is a single number. I have tried doing this by adding a Dense layer with the "units" argument set to 1. However, due to the extra dimensions of my input data, I still end up with an output of shape (None, 13, 8, 1).

import numpy
from keras.layers import Conv2D, Input, Dense
from keras.models import Model
from keras.optimizers import SGD


def make_training_data():
    input_training_data = []
    for i in range(6):
        a = []
        for j in range(13):
            b = []
            for k in range(8):
                c = []
                for l in range(8):
                    c.append(0)
                b.append(c)
            a.append(b)
        input_training_data.append(a)

    output_training_data = []
    for i in range(6):
        output_training_data.append(0)

    return input_training_data, output_training_data


def make_neural_net():
    input = Input((13, 8, 8))
    output = Conv2D(256, (3, 3), input_shape=(13, 8, 8), padding='same')(input)
    output = Dense(1)(output)
    return Model(outputs=output, inputs=input)


def main():
    input_training_data, output_training_data = make_training_data()
    neural_net = make_neural_net()
    input_training_data = numpy.array(input_training_data)
    output_training_data = numpy.array(output_training_data)
    sgd = SGD(0.2, 0.9)
    neural_net.compile(sgd, 'categorical_crossentropy', metrics=['accuracy'])
    neural_net.fit(input_training_data, output_training_data, epochs=1)


main()

How can I change my network so that I get the right size of output?

Upvotes: 1

Views: 421

Answers (1)

Nicolas Gervais
Nicolas Gervais

Reputation: 36584

You need to flatten your tensors before you turn a convolutional layer into a 1D tensor.

Replace:

output = Dense(1)(output)

with:

output = Flatten()(output)
output = Dense(1)(output)`

Import it from keras.layers.Flatten

Also, since you have 1 output neuron, you will need to change your activation function to 'sparse_categorical_crossentropy'

Upvotes: 3

Related Questions