CODEWITHSUNDEEP
pythonkeras
Segmented
Segmented

Reputation: 2044

Python: keras shape mismatch error

I am trying to build a very simple multilayer perceptron (MLP) in keras:

model = Sequential()
model.add(Dense(16, 8, init='uniform', activation='tanh'))
model.add(Dropout(0.5))
model.add(Dense(8, 2, init='uniform', activation='tanh'))

sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='mean_squared_error', optimizer=sgd)

model.fit(X_train, y_train, nb_epoch=1000, batch_size=50)
score = model.evaluate(X_test, y_test, batch_size=50)

My training data shape: X_train.shape gives (34180, 16)

The labels belong to binary class with shape: y_train.shape gives (34180,)

So my keras code should produce the network with following connection: 16x8 => 8x2

which produces the shape mismatch error:

ValueError: Input dimension mis-match. (input[0].shape[1] = 2, input[1].shape[1] = 1)

Apply node that caused the error: Elemwise{sub,no_inplace}(Elemwise{Composite{tanh((i0 + i1))}}[(0, 0)].0, <TensorType(float64, matrix)>)
Inputs types: [TensorType(float64, matrix), TensorType(float64, matrix)]
Inputs shapes: [(50, 2), (50, 1)]
Inputs strides: [(16, 8), (8, 8)]

At Epoch 0 at line model.fit(X_train, y_train, nb_epoch=1000, batch_size=50). Am I overseeing something obvious in Keras?

EDIT: I have gone through the question here but does not solve my problem

Upvotes: 3

Views: 7310

Answers (1)

mattdns
mattdns

Reputation: 894

I had the same problem and then found this thread;

https://github.com/fchollet/keras/issues/68

It appears for you to state a final output layer of 2 or for any number of categories the labels need to be of a categorical type where essentially this is a binary vector for each observation e.g a 3 class output vector [0,2,1,0,1,0] becomes [[1,0,0],[0,0,1],[0,1,0],[1,0,0],[0,1,0],[1,0,0]].

The np_utils.to_categorical function solved this for me;

from keras.utils import np_utils, generic_utils

y_train, y_test = [np_utils.to_categorical(x) for x in (y_train, y_test)]

Upvotes: 10

Related Questions