I am bulding a multi-task model with two outputs. I saw many coding examples in which every output layer should be given a name that is the same as the one passed for compile function: class Model_3(Model): def __init__(self, data_format='channels_first', **kwargs ): super(Model_3, self).__init__(**kwargs) # ...... Initializing the models block.......... # showing only the two 1x1 conv responsible for generating the 2 outputs: # 1x1 conv. blocks self.conv_1 = Conv2D(filters=1, kernel_size=1, strides=1, padding='same', data_format=self.data_format, name = 'bone_mask') self.conv_2 = Conv2D(filters=1, kernel_size=1, strides=1, padding='same', data_format=self.data_format, name = 'lesion_mask') def build(self, input_shape): super(Model_3, self).build(input_shape) def call(self, inputs): bone_mask = self.conv_1(x_decoder_4) lesion_mask = self.conv_2(x_decoder_8) return{'bone_mask': bone_mask, 'lesion_mask': lesion_mask} model.compile(optimizer=adam, loss={'bone_mask': soft_dice_loss, 'lesion_mask': soft_dice_loss} , loss_weights={ 'bone_mask': task_1_weight, 'lesion_mask': task_2_weight} ) I noticed that if I omit the name argument in the two 1x1 conv layers but if still the model returns a dictionary, tensorflow can still train the model using model.fit : class Model_3(Model): def __init__(self, data_format='channels_first', **kwargs ): super(Model_3, self).__init__(**kwargs) # ...... Initializing the models block.......... # showing only the two 1x1 conv responsible for generating the 2 outputs: # 1x1 conv. blocks self.conv_1 = Conv2D(filters=1, kernel_size=1, strides=1, padding='same', data_format=self.data_format) self.conv_2 = Conv2D(filters=1, kernel_size=1, strides=1, padding='same', data_format=self.data_format) def build(self, input_shape): super(Model_3, self).build(input_shape) def call(self, inputs): bone_mask = self.conv_1(x_decoder_4) lesion_mask = self.conv_2(x_decoder_8) return{'bone_mask': bone_mask, 'lesion_mask': lesion_mask} model.compile(optimizer=adam, loss={'bone_mask': soft_dice_loss, 'lesion_mask': soft_dice_loss} , loss_weights={ 'bone_mask': task_1_weight, 'lesion_mask': task_2_weight} ) So, my question is if I omit the names of the last layers but return a dictionary with keys that match the names specified in the model.compile function (e.g., {'bone_mask': output1, 'lesion_mask': output2}), will TensorFlow still be able to correctly map the loss functions and metrics to the corresponding outputs based solely on the dictionary keys? I'm asking because I've already done a significant amount of work without explicitly naming the last layers, and I want to ensure that this omission won't cause any issues. Continue reading...
