Reshaping layer in Keras

Question

I am trying to build a convolutional neural network with an output matrix. The input shape is (100,100,4) and the output shape is (2,125).

Here is the summary of my current model:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_63 (InputLayer)        (None, 100, 100, 4)       0         
_________________________________________________________________
conv2d_44 (Conv2D)           (None, 100, 100, 25)      2525      
_________________________________________________________________
max_pooling2d_38 (MaxPooling (None, 50, 50, 25)        0         
_________________________________________________________________
flatten_38 (Flatten)         (None, 62500)             0         
_________________________________________________________________
dense_47 (Dense)             (None, 10)                625010    
_________________________________________________________________
dense_48 (Dense)             (None, 250)               2750      
_________________________________________________________________
reshape_63 (Reshape)         (None, 2, 125)            0         
=================================================================
Total params: 630,285
Trainable params: 630,285
Non-trainable params: 0
_________________________________________________________________
None

Which I thought should be fine but when I tried to fit the model I got this error:

ValueError: Error when checking target: expected reshape_62 to have shape (2, 1) but got array with shape (2, 125)

Here's the code I used

batch_size = 100

input_layer     = Input(shape=(xs[1],xs[2],xs[3]))
conv1           = Conv2D(filters = 25, kernel_size = 5,padding="same",activation="relu", data_format = 'channels_last')(input_layer)
pool1           = MaxPooling2D(pool_size=(2,2),padding="same")(conv1)
flat            = Flatten()(pool1)
hidden1         = Dense(10, activation='relu')(flat)
output_layer    = Dense(ys[1]*ys[2], activation='softmax')(hidden1)
output_reshape  = Reshape((2,125))(output_layer)
model           = Model(inputs=input_layer, outputs=output_reshape)
print(model.summary())

model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', sample_weight_mode='temporal')
model.fit(x_train,y_train,batch_size=batch_size,epochs=3)

I have been looking up how the reshape layer work but I still couldn't figure it out. Any help would be most appreciated.

Answer 1

This is happening because you are using 'sparse_categorical_crossentropy'.

"Sparse" means that the system will not expect an entire array, but just the cordinate of the hot spot. Instead of expecting a regular (None, 2, 125) tensor, it will expect just (None, 2, 1) indicating which of the 125 classes is the correct one.

To fix this, either you start using a sparse y_train, or you replace your loss with 'categorical_crossentropy'.

I believe a sparse y_train can be obtained with sparse_y_train = numpy.argmax(y_train, axis=-1). If this model does not gives you memory problems, you don't need to go sparse.