You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
ONNX models used in practice often violate some of the contraints specified by the ONNX standard. Examples include the following:
The standard specifies that identifiers must be valid C identifiers, but use of numbers such as "23" as tensor-names is common.
ONNX has a more complex shape-specification that uses symbolic identifiers to denote unknown dimensions, but occasionally models use a number such as -1 to indicate an unknown dimension (common in some other frameworks).
A sanitizer tool that fixes such violations (where a simple fix exists) would be useful. It would take in an input ONNX model and produce a sanitized ONNX model as output with such obvious violations repaired.
The text was updated successfully, but these errors were encountered:
I have seen exported models where providing values for the exported symbols do not work because constants are used in some places and the symbol in others. We let shapes be provided for inputs and warn/err as appropriate if shape inference finds inconsistencies.
ONNX models used in practice often violate some of the contraints specified by the ONNX standard. Examples include the following:
A sanitizer tool that fixes such violations (where a simple fix exists) would be useful. It would take in an input ONNX model and produce a sanitized ONNX model as output with such obvious violations repaired.
The text was updated successfully, but these errors were encountered: