Multidimensional Input to Keras

Question

I am new to keras and I am having a problem running a model with more than one dimension. So, I have been trying a few samples. This is one of them.

from keras.models import Sequential
from keras.layers import Dense
import numpy as np

X_train = np.array([[[1, 2], [3, 4]], [[1, 2], [3, 4]]])
model = Sequential([
    Dense(32, input_shape=X_train.shape[1:]),
])
model.compile(loss='sparse_categorical_crossentropy', optimizer='sgd')
model.fit(X_train, [1, 2])

I expect the above sample to run but I am getting an error

Error when checking target: expected dense_1 to have 3 dimensions, but got array with shape (2, 1)

What can be the reason for this. Can anyone give an example on how to run keras model with a multi-dimensional input, i.e. how should the input be structured? Thank you.

Answer 1

I would suggest that you one-hot encoded your output classes, i.e. using:

# Convert labels to categorical one-hot encoding
labels = np.array([1, 2]) # 0 - num_classes - 1
y_train = keras.utils.to_categorical(labels, num_classes=3)

And then used 'categorical_crossentropy' instead of 'sparse_categorical_crossentropy':

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

Finally you should flatten your 3D-input at some point if you want 2D-output (samples x classes). Number of units in the output layer (you have only one) should match number of classes and use an appropriate activation function (e.g. 'softmax')

model.add(Flatten(input_shape=X_train.shape[1:]))
model.add(Dense(3, activation='softmax'))

Try to have a look at the Multilayer Perceptron (MLP) for multi-class softmax classification at: https://keras.io/getting-started/sequential-model-guide/