Error when checking target: expected dense_3 to have shape (3,) but got array with shape (1,)

Question

I am working on training a VGG16-like model in Keras, on a 3 classes subset from Places205, and encountered the following error:

ValueError: Error when chec         


        

Answer 1

The reason for this is you would have used 'binary' class_mode in the fit_generator() method for a multi class problem. Change that to 'categorical' and the error goes.

Answer 2

Had the same issue. To solve the problem you can simply change in validation_generator and train_generator the class mode from 'binary' to 'categorical' - that's because you have 3 classes-which is not binary.

Answer 3

I also got the same error and solved it by setting class_mode as categorical instead of binary

Answer 4

The problem is with the shape of the labels of the data "Y".
The shape you have for the labels are (m,) and this will not work with the:

loss = "binary_crossentropy"

I believe if you don't want to play with the shape of the labels, then use:

loss = "sparse_categorical_crossentropy"

Answer 5

As mentioned by others, Keras expects "one hot" encoding in multiclass problems.

Keras comes with a handy function to recode labels:

print(train_labels)
[1. 2. 2. ... 1. 0. 2.]

print(train_labels.shape)
(2000,)

Recode labels using to_categorical to get the correct shape of inputs:

from keras.utils import to_categorical
train_labels = to_categorical(train_labels)

print(train_labels)
[[0. 1. 0.]
 [0. 0. 1.]
 [0. 0. 1.]
 ...
 [0. 1. 0.]
 [1. 0. 0.]
 [0. 0. 1.]]

print(train_labels.shape)
(2000, 3)  # viz. 2000 observations, 3 labels as 'one hot'

Other importent things to change/check in multiclass (compared to binary classification):

Set class_mode='categorical' in the generator() function(s).

Don't forget that the last dense layer must specify the number of labels (or classes):

model.add(layers.Dense(3, activation='softmax'))

Make sure that activation= and loss= is chosen so to suit multiclass problems, usually this means activation='softmax' and loss='categorical_crossentropy'.

Answer 6

For me, this worked.

from keras.utils import to_categorical

num_labels=10 #for my case

train_labels=to_categorical(train_labels,10)
test_labels=to_categorical(test_labels,10)

Specifying the number of labels as an argument while categorically encoding my labels helped train effectively on my training set.