[FIXED] Exporting ONNX Model from Deep Learning Frameworks at Operator-level

Issue

Hi, I have some questions regarding exporting ONNX models.

Let's say we have an LSTM cell from PyTorch.

Using torch.onnx.export produces ONNX model with LSTM layer.

However, I am interested in whether it can produce the ONNX model at the operator level, i.e, matmul, add.

Is there a way to do so?

If not, is there another way to make an operator level ONNX model?

Thanks, Jake

Solution

When you export a model from PyTorch to onnx using the torch.onnx.export() function, it records all the operations that the initial model has used. as mentioned here.

we call the torch.onnx.export() function. This will execute the model, recording a trace of what operators are used to computing the outputs.

so, yes it does produce the onnx model at the operator level, you can even visualize the exported .onnx model graph using netron

if you still want to use the onnx operator, here is the ONNX Operator Schemas.

Answered By - kiranr