Keras Failed to convert LSTM layers to ONNX on Torch backend

Not sure if this is Torch or Keras issue, but simple LSTM models cannot be exported as ONNX.

TypeError: Exception encountered when calling LSTM.call().

unsupported operand type(s) for -: 'torch.return_types.max' and 'int'

Arguments received by LSTM.call():
  • sequences=torch.Tensor(shape=torch.Size([1, 32, 256]), dtype=float32)
  • initial_state=None
  • mask=None
  • training=None

Here's the link to colab notebook reproducing the issue: https://colab.research.google.com/drive/1QCTYR9xCConwt4D4R3rXUxvAeO0Gd2IG

P.S. The same model will convert fine with TF backend. P.P.S. The verbosity flag is not respected.

Comment From: dhantule

Hi @DLumi, thanks for reporting this.

I have run your code and I'm facing the same error in this gist. We'll look into this and update you.

Comment From: mattdangerw

Thanks for filing. This does indeed look like a bug.

@james77777778 any thoughts on this?

I'll also mark this open for contributions in case anyone would like to take this on. Please chime in on the bug if you do.