We require users to manually specifying the number of inputs and outputs for layers in neural network protocol buffer configuration files. However, the number of inputs can be determined by the number of outputs of the previous layer or input_shape (for the first layer), and the number of outputs can be determined by the number of inputs or layer-specific configurations (e.g. for activation layers, num_output is equal to num_input and for fully connected layers, num_output needs to be specified by the configuration). Thus, we can infer the number of inputs and outputs of layers without users' manual configurations.
Also, these manual configurations can cause inconsistency. For example, in the configuration for activation layers, users may set different values for num_input and num_output. To prevent this kind of situation, we should remove required num_input and num_output fields.
We require users to manually specifying the number of inputs and outputs for layers in neural network protocol buffer configuration files. However, the number of inputs can be determined by the number of outputs of the previous layer or
input_shape
(for the first layer), and the number of outputs can be determined by the number of inputs or layer-specific configurations (e.g. for activation layers,num_output
is equal tonum_input
and for fully connected layers,num_output
needs to be specified by the configuration). Thus, we can infer the number of inputs and outputs of layers without users' manual configurations. Also, these manual configurations can cause inconsistency. For example, in the configuration for activation layers, users may set different values fornum_input
andnum_output
. To prevent this kind of situation, we should remove requirednum_input
andnum_output
fields.