Validate that raw_data XOR floats/etc field is used in TensorProto about onnx

[bddppq]

For each data type the data can be stored in the corresponding data field (e.g. "FLOAT" -> float_data) or raw_data (except string, which can only be stored in "string_data" field). Storing it in raw_data field is for efficient protobuf parsing, see here: https://github.com/onnx/onnx/blob/master/onnx/numpy_helper.py#L46

[ezyang]

The rule is that either raw_data is set, OR one of the other fields is set, appropriate for the type. @bddppq, do you know if the validator is testing for this? It's an error for multiple to be simultaneously set. Maybe we should have just gotten rid of the non-data fields.

In any case, float data will only ever be in float_data or raw_data, you don't have to test for ints.

[bddppq]

I don't think there is any place checking this, we could check initializer in check_graph.

[ezyang]

Yep; also we need to check it for tensor constants

[linkerzhang]

Thanks! We may also come out of helper functions to parse the raw_data in this way. :).

[ezyang]

Keeping this issue open to track validating tensor proto format.

[ezyang]

Aaand fixed now :)