onnx2pytorch icon indicating copy to clipboard operation
onnx2pytorch copied to clipboard

Missing type STRINGS when parsing attributes

Open q-ycong-p opened this issue 3 years ago • 2 comments

The current implementation doesn't support STRINGS as attribute type. Per ONNX sepc for LSTM, one of the attributes can be activations which is of type STRINGS. This leads to conversion error on LSTM models. Could you add STRINGS attribute type?

q-ycong-p avatar Jan 26 '22 00:01 q-ycong-p

Hi @calvinmccarter-at-lightmatter and @Talmaj ,

As a follow-up to the original request, I understand that STRINGS attrib being missing might be related to the fact that activation attribute is not current supported (refers here). If possible, can we please add support for explicitly listed torch-compatible activations? Since some 3p-converter-generated onnx models by default specify the activation fields (e.g. ['Sigmoid','Tanh','Tanh'] for LSTM).

q-ycong-p avatar Jan 27 '22 00:01 q-ycong-p

Attribute STRINGS added in v0.5.2, but I think the activations in LSTM or GRU are still not supported.

Talmaj avatar Nov 01 '25 23:11 Talmaj