Bug Report

Is the issue related to model conversion?

No.

Describe the bug

The parser has only partial support for data type parsing:

The missing types are:

TensorProto_DataType_FLOAT16
TensorProto_DataType_BFLOAT16
TensorProto_DataType_FLOAT8E4M3FN
TensorProto_DataType_FLOAT8E4M3FNUZ
TensorProto_DataType_FLOAT8E5M2
TensorProto_DataType_FLOAT8E5M2FNUZ
TensorProto_DataType_COMPLEX64
TensorProto_DataType_COMPLEX128

System information

System independent.

Reproduction instructions

Find a sample model and instruction for float16 below; note however that it would be great to support all types not just float16. The error will be the same for the other types.

Sample model:
gemmfloat16.onnxtxt

Sample instructions:

import onnx
from pathlib import Path
onnx.parser.parse_model(Path("gemmfloat16.onnxtxt.txt").read_text())

Result:
onnx.parser.ParseError: b'[ParseError at position (...)]\nError context: <float16[4, 4] weight = {...}, float16[4] bias = {...}>\nUnhandled type: %d10'

(Note: The .onnxtxt format is not allowed to be uploaded therefore the sample is .onnxtxt.txt)

Expected behavior

ONNX model is parsed successfully.