Version confusion when ONNC compile ONNX models
wang-y-z opened this issue · 0 comments
Background
When I compile shufflenetv2.onnx model(format ONNX v4/ torch1.2/ ai.onnx v8) in the newest docker supported by ONNC (1.3.0), come out the Fatal message:
Fatal:Cannot inference shape of this onnx model, got exception from ONNX: Your model ir_version is higher than the checker's
And I check /models/ in ONNC docker（onnc-community:1.3.0）
all the models built by:
format ONNX v3 producer pytorch 1.2 imports ai.onnx v8
My questions:
1: How to export onnx model with (v3, 1.2, v8) by pytorch ?
(1) I have tried a lot and get this result:(use torch.onnx.export)
| pytorch | format ONNX ir_version | producer pytorch | imports ai.onnx |
|---|---|---|---|
| 1.4 | v4 | 1.3 | v8~ |
| 1.2 | v4 | - | v8~ |
| 1.1 | v4 | - | v9~ |
| 1.0 | v3 | 0.4 | v9~ |
| 0.4 | v3 | 0.3 | v6~ |
It seems that ONNC 1.3.0 only support ir_version v3.
So how to export onnx v3 with producer 1.2 ? torch.onnx.export-source code
Can you share the script which is to export ONNX(v3,1.2,v8) or can you package the environment into a docker and share it please?
(2) The new ONNX model has achieved v6, can anybody share the develop plan for the ir_version please?