CODEWITHSUNDEEP
pythonkeras
Rehaan Ahmad
Rehaan Ahmad

Reputation: 814

Adding Custom Root Mean Square Error In Keras

I am trying to write a RMSE function in Keras that only runs the RMSE over array values that are not zero. I have two arrays arr1 and arr2. Both arrays have zeros in exactly the same places (thus they contribute zero to the RMSE value). However, I need to change the number I am dividing by to the number of non zero values in arr1 (or arr2)

def root_mean_squared_error(y_true, y_pred):
    nonzero = tf.count_nonzero(y_pred)
   num_zeros=tf.reduce_sum(tf.where(tf.not_equal(y_pred,0),tf.ones_like(y_pred),tf.zeros_like(y_pred))) 
    return K.sqrt((K.sum(K.square(y_pred - y_true))/tf.cast(nonzero, tf.float32)))

mc = keras.callbacks.ModelCheckpoint('modelsPerEpoch/weights{epoch:06d}.hdf5', 
                                     save_weights_only=False, 
                                     period=1)

decay_learner = ValidationLearningRateScheduler()

main_input = Input(shape=(None, 2, 100, 100), dtype='float32', name='input')

mask=Input(shape=(1, 100, 100), dtype='float32', name='mask')

hidden = ConvLSTM2D(filters=16, 
                    kernel_size=(5, 5),  
                    padding='same',  
                    return_sequences=False, 
                    data_format='channels_first')(main_input)

output = Conv2D(filters=1, 
                kernel_size=(1, 1), 
                padding='same',
                activation='sigmoid',
                kernel_initializer='glorot_uniform',
                data_format='channels_first',
                name='output')(hidden)

output_with_mask=Multiply()([output, mask])

sgd = SGD(lr=0.002, momentum=0.0, decay=0.0, nesterov=False)

model = Model(inputs=[main_input, mask], outputs=output_with_mask)

model.compile(optimizer=sgd,
              loss=root_mean_squared_error,
              metrics=[metrics.mse, root_mean_squared_error])

However, when I run this, I get an "inf" returned in the command line. How can I fix this?

Upvotes: 1

Views: 1083

Answers (1)

giser_yugang
giser_yugang

Reputation: 6166

y_true and y_pred have zeros in exactly the same places is not valid according to your code. You get inf in the command line because the non-zero number in y_pred is 0, that is nonzero = 0 in your code. You should count the correct non-zero numbers and avoid dividing by 0 by the following code.

def root_mean_squared_error(y_true, y_pred):
    nonzero = tf.count_nonzero(y_true)
    ...
    return K.switch(K.equal(nonzero,0)
                    , K.constant(value=0.)
                    , K.sqrt((K.sum(K.square(y_pred - y_true))/tf.cast(nonzero, tf.float32))))

Upvotes: 1

Related Questions