Keras ValueError: Error when checking target: expected dense_1 to have 3 dimensions

Question

I'm using keras with tensorflow backend & have faced problem figuring out the right shapes for layers for my model.

I have already read this useful explanation on the difference of various keras layers' attributes.

This is the architecture of my model:

I'm trying to do a binary classification (logistic regression) using categorical labels & hence the last layer is a Dense layer with 1 unit which I assumed that would evaluate to 1 for positive class and 0 for negative class.

And this is my model's summary:

My input in one side of the net is 10158 & for the other-side is 20316. I have total 1370 samples. My train_data's shape is (1370, 1, 10158) & the label's shape is (1, 1370) & the batch size is 100.

input_layer = Input(shape=(1,no_terms), name='docs')
s = Lambda(lambda x: x+1)(input_layer)
log_layer = Lambda(log, name='tf_output')(input_layer)

tpr_fpr = np.zeros((2, no_terms))
tpr_fpr[0,:] = np.sum(train_docs[np.where(train_label>0), :]>0, axis=1
                      )/np.sum(train_label>0) * (1000)
tpr_fpr[1,:] = np.sum(train_docs[np.where(train_label>0), :]>0, axis=1
                     )/np.sum(train_label <= 0) * (1000)

k_constants = backend.constant(np.reshape(tpr_fpr.T, (1,2*no_terms)))
fixed_input = Input(tensor=k_constants, shape=(1, 2*no_terms), name='tpr_fpr')
h = Dense(int(300), activation='relu', name='hidden', input_shape=(1, 2*no_terms), 
          trainable=True)(fixed_input)
h = Dropout(0.2, name="D")(h)
cd = Dense(units=no_terms, activation='relu', name='cd', trainable=True)(h)


prod = Multiply()([log_layer, cd])
o = Lambda(lambda x:(x/backend.sqrt(backend.sum(x * x,axis=1,keepdims=True))))(prod)
o = ReLU(max_value=None, negative_slope=0.0, threshold=0.0)(o)
o = Dense(1, activation='sigmoid', input_shape=(no_terms,))(o)


model_const = Model(fixed_input,cd)
model = Model([input_layer, fixed_input], o)

op = optimizers.RMSprop(learning_rate=.1, rho=0.9)
model.compile(optimizer=op, loss=mean_squared_error, metrics=['accuracy'])
plot_model(model, to_file='model.png')
model.summary()
batchSize = 100

checkpoint = ModelCheckpoint(filepath="a.hdf5",monitor='val_acc', mode='max', 
                             save_best_only=True)
earlystop=EarlyStopping(monitor='val_loss', patience=20)

train_docs.shape = (train_docs.shape[0], 1, train_docs.shape[1])
train_label = to_categorical(train_label, num_classes=2, dtype='float32')
model.fit(train_docs, train_label, epochs=10, batch_size=batchSize, 
          validation_data=(test_docs, test_label), 
          callbacks=[earlystop, checkpoint], verbose=1)

And here's the error I got: "ValueError: Error when checking target: expected dense_1 to have 3 dimensions, but got array with shape (1430, 2)"

I have no Idea what is the what the (1430, 2) shape refers to & why I got this error.

Answer 1

You did indeed directly solve the problem - here's how:

• Keras binary classification expects labels ('targets') shaped as (batch_size, 1). Reason: the goal of the final layer is to output predictions, which will be compared against labels to compute metrics (loss, accuracy, etc) - and labels are shaped (batch_size, 1)
• Above is also why to_categorical was a problem - see snippet from docs below; for binary classification, one-hot encoding is redundant, as binary_crossentropy directly compares labels against predictions as supplied
• Keras Dense expects inputs to be 2D: (batch_size, input_dim). Your reshaping was making the input 3D: (batch_size, 1, input_dim)
• Above is also why shape=(1, no_terms) --> shape=(no_terms,) helped; both are, in fact, correct for the data shapes you were feeding at the time. The full batch shape simply includes the batch dim: (batch_size, no_terms) (no_terms == input_dim)
• Lastly, for binary classification, use loss='binary_crossentropy' - and never mean square error for classification problems (unless for very specific reasons)

# Consider an array of 5 labels out of a set of 3 classes {0, 1, 2}:
> labels
array([0, 2, 1, 2, 0])
# `to_categorical` converts this into a matrix with as many
# columns as there are classes. The number of rows
# stays the same.
> to_categorical(labels)
array([[ 1.,  0.,  0.],
       [ 0.,  0.,  1.],
       [ 0.,  1.,  0.],
       [ 0.,  0.,  1.],
       [ 1.,  0.,  0.]], dtype=float32)

Answer 2

Well I found the solution to the error but still can't understand why that previous shape didn't work out & quite frankly this fix was a success among numerous trial & error.

I changed the following layers input from the shape format (1, x) to the format of (x,):

    input_layer = Input(shape=(no_terms,), name='docs')

    k_constants = backend.constant(np.reshape(tpr_fpr.T, (1,2*no_terms)))
    fixed_input = Input(tensor=k_constants, shape=(2*no_terms,), name='tpr_fpr')
    h = Dense(int(300), activation='relu', name='hidden', input_shape=(2*no_terms,), trainable=True)(fixed_input)

    o = ReLU(max_value=None, negative_slope=0.0, threshold=0.0)(o)
    o = Dense(1, activation='sigmoid', input_shape=(no_terms,))(o)

and also removed the following lines from the code:

    train_docs.shape = (train_docs.shape[0], 1, train_docs.shape[1])
    train_label = to_categorical(train_label, num_classes=2, dtype='float32')

Now I'm just using a label of shape (#no_of_samples, 1) which is binary & not a categorical label.

So the new structure is:

I hope somebody can explain what was wrong with the previous model, so I would avoid making the same mistake again.

Thanks.

Answer 3

checking target: expected dense_1 to have 3 dimensions, but got array with shape (1430, 2)"

That means dense_1 have 3 dimensions but your input have only 2 dimensions if your design this model in Image processing in that case you have declare the image shape yhat is ((48,48),1) & ((48,48),3) here 1 is for gray scale & 3 is for rgb images