Do layer names get imported into Relay?

Questions
1

Hello,

In some frameworks (maybe all?) we can define a layer/operator name. Looking at the Relay Python API documentation, I dont see that Relay operators have a ‘name’ (except Bitpack which uses this parameter for a type of bitpacking). Also when I import a model from TF and inspect the Relay representation, I don’t see any indication that the names of the operators are available at the Relay representation.

Am I missing something or are the names really not available at the Relay level?

Thanks

1 Like
2

Previous IR (NNVM) has one name attribute but relay miss it.

3

Thanks for the reply.

I noticed that at least the inputs, weights and biases get the name <framewok_layer_name>/<name_of_type_of_arg> (so for example ‘conv2d_1/input’ would be the IFM of the ‘conv2d_1’ layer from the model). So I guess over the naming of the arguments its possible to determine some relationships between relay operators and the framework layers.

Do you have an intuition why the ‘name’ attribute was dropped? and an idea how much effort it would be to include it back again?

4

I don’t know why relay drop it, but i think @tqchen @jroesch maybe know more information and background.