how to design a custom metric in keras use callbacks argument of model.fit?

Question

I want to design a metric to estimate the extent to which the two distributions generated from a Siamese model overlap during training, but don't find any solution.

I have referenced this discussion to construct the metric using callbacks argument of model.fit:

class Metrics(tf.keras.callbacks.Callback):
    def on_train_begin(self, logs={}):
        self._data = []

    def on_epoch_end(self, batch, logs={}):
        x1, x2 = self.validation_data[0], self.validation_data[1]
        predictions = np.asarray(siamese.predict([x1, x2]))
        
        data_distribution1 = predictions[np.squeeze(np.where(labels_data[train_index] == 1))]
        data_distribution2 = predictions[np.squeeze(np.where(labels_data[train_index] != 1))] 

        d1=list(data_distribution1)
        d2=list(data_distribution2)
        data_list = [d1, d2]
        robjects.globalenv['data_list'] = robjects.ListVector({f'data_vector{i+1}': robjects.FloatVector(data_vector) for i, data_vector in enumerate(data_list)})
        overlap_coef = robjects.r('overlap(data_list)')
        #print('overlap:',overlap_coef[0][0])

        # y_val = np.argmax(y_val, axis=1)
        overlap_indx = np.argmax(overlap_coef[0][0], axis=1)

        self._data.append({
            'overlap': overlap_indx,
        })
        return

    def get_data(self):
        return self._data

def create_model(distance_metric):
input = layers.Input((photo_size, photo_size, channel))
x = layers.Conv2D(32, (3, 3), activity_regularizer=l2(0.001), activation = 'relu')(input)
x = layers.MaxPooling2D(pool_size=(2, 2))(x)
x = layers.Dropout(0.7)(x)
x = layers.Conv2D(64, (3, 3), activation = 'relu',padding='same')(x)
x = layers.MaxPooling2D(pool_size=(2, 2))(x)
x = layers.Dropout(0.7)(x)
x = layers.Conv2D(128, (3, 3), activation = 'relu',padding='same')(x)
x = layers.MaxPooling2D(pool_size=(2, 2))(x)
x = layers.Dropout(0.7)(x)
x = layers.Flatten()(x)

    x = layers.Dense(512, activation = 'relu', kernel_initializer='glorot_uniform')(x)
    embedding_network = keras.Model(input, x)
    
    input_1 = layers.Input((photo_size, photo_size, channel))
    input_2 = layers.Input((photo_size, photo_size, channel))
    
    tower_1 = embedding_network(input_1)
    tower_2 = embedding_network(input_2)
    
    if distance_metric == 'cosine':
        merge_layer = layers.Lambda(cosine_similarity)([tower_1, tower_2])
    else:
        merge_layer = layers.Lambda(euclidean_distance)([tower_1, tower_2])
    normal_layer = layers.BatchNormalization()(merge_layer)
    model = keras.Model(inputs=[input_1, input_2], outputs=normal_layer)
    return  model

distance = ['euclidean']
siamese = create_model(distance)
siamese.compile(loss=loss, optimizer="SGD", metrics=\["accuracy"\])

metrics = Metrics()
history = siamese.fit(
    [x_train_1, x_train_2],
    y_train,
    batch_size=batch_size,
    epochs=epochs,
    validation_data=(x_train_1, x_train_2),
    callbacks=[metrics]
    )
metrics.get_data()

Here are the problems I encountered:

1. It does not seem to work, as I got this error message "AssertionError: Could not compute output Tensor("batch_normalization/batchnorm/add_1:0", shape=(None, 1), dtype=float32)"
2. Is it possible to directly read the data [x_train_1, x_train_2]] in the Metrci class?

Many thanks!