This is in regards to the PR that adds LSTM support for ONNX parser - https://github.com/apache/incubator-tvm/pull/4825/files
Currently, the code assumes that the input names of the ONNX LSTM ops are same as listed in at https://github.com/onnx/onnx/blob/master/docs/Operators.md#inputs-3---8
The code in Relay parser is
However, I have a model whose input names are not exactly these
So, IIUC, we might have misread the ONNX spec. Basically, we should use onnx_input length to understand how many attributes are optional. And then we should use inputs[index_number] directly to parse instead of inputs[“initial_c”] etc.
@jwfromm I have 1 day ONNX experience, not sure if I understood correctly. Let me know what you think here.
cc @adb