[optimizer] Allow optimizer to function when external data is not available

[justinchuby]

Models with external data may not have their data available when being passed into the optimizer. We should still optimize what we can without raising an error.

[justinchuby]

cc @gramalingam

[gramalingam]

I took a quick look at the IR-based constant-folder/optimizer. I think it should work, as long as the numpy() method returns None or value.const_value is None. We can change it to use whatever appropriate methods the IR exposes to access such external tensors, allowing for possibility that the value may not be available.

On a different note: I just remembered that ir.Value.const_value is used for initializers that are graph inputs as well. This could break a few things. The optimizer etc. assume .const_value has a value only for constants, but graph inputs that are initializers should NOT be treated as constants.

[gramalingam]

Is this manifesting itself in any benchmark run (or another model)? Perhaps for proto-based optimization? We should deprecate that anyway.

[justinchuby]

I took a quick look at the IR-based constant-folder/optimizer. I think it should work, as long as the numpy() method returns None or value.const_value is None. We can change it to use whatever appropriate methods the IR exposes to access such external tensors, allowing for possibility that the value may not be available.

On a different note: I just remembered that ir.Value.const_value is used for initializers that are graph inputs as well. This could break a few things. The optimizer etc. assume .const_value has a value only for constants, but graph inputs that are initializers should NOT be treated as constants.

Right. We shouldn’t assume only Constants have const_value. Any value can have constant values.