Keras Train Neural Network Dimension Value Error: expected to have 2 dimensions, but got array with shape (32, 1, 4)

Question

Python 3.6
Keras 2.2
Tensorflow 1.8 backend

I am having trouble training my neural network because I am getting this error:

ValueError: Error when checking target: expected t_dense_3 to have 2 dimensions, but got array with shape (32, 1, 4)

My neural network

>>> sgd = optimizers.SGD(lr=0.01, decay=1e-6)
>>> target_q_network = Sequential([
      Dense(40, input_shape=observation_shape, activation='relu', name='t_dense_1'),
      Dense(40, activation='relu', name='t_dense_2'),
      Dense(number_of_actions, activation='linear', name='t_dense_3')
    ])
>>> target_q_network.compile(loss='mean_squared_error', optimizer=sgd)
>>> observation_shape
    (8,)

-----------------------------------------------------------------

(Pdb) target_q_network.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
t_dense_1 (Dense)            (None, 40)                360       
_________________________________________________________________
t_dense_2 (Dense)            (None, 40)                1640      
_________________________________________________________________
t_dense_3 (Dense)            (None, 4)                 164       
=================================================================
Total params: 2,164
Trainable params: 2,164
Non-trainable params: 0
_________________________________________________________________

When I pass in values to the neural network an array of shape (1, 4) is returned:

(Pdb) env.reset()
array([-0.00126171,  0.94592496, -0.12780861,  0.35410735,  0.00146875, 0.02895054,  0.        ,  0.        ])
# Passing value into Neural Network
(Pdb) target_q_network.predict(env.reset().reshape(1,8))
array([[ 0.07440183,  0.03480911,  0.11266299, -0.08043154]], dtype=float32)

I am passing in training_set and labels

(Pdb) training_set.shape
(32, 8)
(Pdb) labels.shape
(32, 1, 4)

Answer 1

The 'mean_squared_error' loss function is probably expecting to receive a (batch_sz x n_labels) matrix of labels, but you are passing a (batch_sz x 1 x n_labels) matrix of labels, specifically with labels.shape=(32, 1, 4) . You just need to reshape your labels to have the shape (batch_sz x n_labels) so it has labels.shape=(32, 4) , which can then be properly compared to the neural network output.