Convert .Nemo to CoreML

[monologue110]

I've tried to convert both the quartnet.nemo model and citrinet.nemo model output to the coreML format.
However, in the process of torch.jit.trace() , it always shows the argument error. Originally I did torch.jit.trace(model_ev, example_input)

Then I tried to change the args to torch.jit.trace(model_ev, kwarg) where kwarg is {'processed_signal': single_example[0], 'processed_signal_len': single_example[1]} now it shows AttributeError: 'EncDecCTCModelBPE' object has no attribute 'processed_signal'

Any idea how to really convert the model to coreML?

I've also tried to convert ONNX to CoreML and have not been able to complete it.

---

I've tried to fix it, but during .convert() ,
I got an error message RuntimeError: PyTorch convert function for op 'stft' not implemented.

Even when I change the torch.jit.trace() to torch.jit.script(), I got the same error message :

RuntimeError:
Unknown type name 'STFT':
File "/home/jovyan/.local/lib/python3.7/site-packages/nemo/collections/asr/parts/preprocessing/features.py", line 375
def get_seq_len(self, seq_len):
if isinstance(self.stft, STFT): ~~~~ <--- HERE
pad_amount = self.stft.pad_amount * 2
else:
'FilterbankFeatures.get_seq_len' is being compiled since it was called from 'FilterbankFeatures.forward'
File "/opt/conda/lib/python3.7/site-packages/torch/autograd/grad_mode.py", line 388
def forward(self, x, seq_len):
seq_len = self.get_seq_len(seq_len.float())

[titu1994]

We don't support export of the preprocessor. You can use model.export("filename.ts") to export to torchscript but we don't expect it to work on CoreML since we don't test it for such use c