Error while exporting model after fine tuning

[black-born]

Hello everyone,

After some struggle I finally succeded to launch a fine tuning.
But when I tried to export it to onxx I got this error:

The following call raised this error:
File "/home/jupyter/piper1-gpl/src/piper/train/vits/transforms.py", line 176, in rational_quadratic_spline
assert (discriminant >= 0).all(), discriminant

python3 -m piper.train.export_onnx --checkpoint lightning_logs/version_3/checkpoints/epoch=4393-step=43560.ckpt --output-file piper_test/fr-piper_test-medium.onnx

I also tried directly from the checkpoint I was fine tuning and got the same error, I'm very confuse

python3 -m piper.train.export_onnx --checkpoint cheepoch3304-step2050940.ckpt --output-file piper_test/fr-piper_test-medium.onnx

Hope you can help me,
Have a great day

[irakhlin]

I am also seeing this issue after fine tuning. Using python 3.13.11

[w5xd]

I get the same "discriminant" assertion and don't need to do any training. I simply did piper.train.export from this checkpoint:
https://huggingface.co/datasets/rhasspy/piper-checkpoints/resolve/main/en/en_US/sam/medium/epoch%3D4688-step%3D106008.ckpt
fails with your same error.

I ran training using that same epoch4688 checkpoint using my own recordings up to epoch 4999 and tried export. Same "discriminant" failure.

[Alimjoo]

I am also having this issue, any solution yet ?

[black-born]

I found a solution I just need time to make a pull request.
You need to add a parameter inside the extraction for it to work
Inside src/piper/train/export_onnx.py you need to change
torch.onnx.export( model=model_g, args=dummy_input, f=output_path, verbose=False, opset_version=OPSET_VERSION, input_names=["input", "input_lengths", "scales", "sid"], output_names=["output"], dynamic_axes={ "input": {0: "batch_size", 1: "phonemes"}, "input_lengths": {0: "batch_size"}, "output": {0: "batch_size", 2: "time"}, }, )
by
torch.onnx.export( model=model_g, args=dummy_input, f=output_path, dynamo=False, verbose=False, opset_version=OPSET_VERSION, input_names=["input", "input_lengths", "scales", "sid"], output_names=["output"], dynamic_axes={ "input": {0: "batch_size", 1: "phonemes"}, "input_lengths": {0: "batch_size"}, "output": {0: "batch_size", 2: "time"}, }, )

Pytorch make some change after 2.6 that are not compatible with the model for onxx export so you need dynamo=False to go back to pre 2.6 change

[Alimjoo]

Thanks, this absolutly works

[w5xd]

#140 (comment)
worked for me. Thanks!