Finetuned S & M model export to ONNX

[Aduomas]

Hello,

I was unable to export finetuned (or any) model to ONNX.

I am using:

• pytorch 2.5.1
• torchaudio 2.5.1
• Python 3.10.10 (tried with 3.12.8 as well)

Code:

import torch

device = torch.device("cpu")

model = torch.hub.load(
    "IDRnD/ReDimNet",
    "ReDimNet",
    model_name="S",
    train_type="ft_mix",
    dataset="vb2+vox2+cnc",
).to(device)
model.eval()

input_sample = torch.randn(1, 32000).to(device)

dynamic_axes = {"feats": {0: "B"}, "embs": {0: "B"}}

torch.onnx.export(
    model,
    input_sample,
    "model_success.onnx",
    opset_version=17,
    input_names=["feats"],
    output_names=["embs"],
    dynamic_axes=dynamic_axes,
    verbose=True,
)

Received error:

Inputs:
        #0: 1060 defined in (%1060 : Float(*, *, *, strides=[16512, 16512, 1], requires_grad=0, device=cpu) = onnx::Pad[mode="reflect"](%1034, %1059), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram # /Users/aduomas/.pyenv/versions/3.10.10/lib/python3.10/site-packages/torch/nn/functional.py:5096:0
    )  (type 'Tensor')
        #1: 1070 defined in (%1070 : int[] = prim::ListConstruct(%1064, %1069), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram
    )  (type 'List[int]')
    Outputs:
        #0: 1071 defined in (%1071 : Float(*, *, strides=[16512, 1], requires_grad=0, device=cpu) = onnx::Reshape[allowzero=0](%1060, %1070), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram # /Users/aduomas/.pyenv/versions/3.10.10/lib/python3.10/site-packages/torch/functional.py:703:0
    )  (type 'Tensor')

[Aduomas]

Google Colab environment doesn't work either.

I was also unsucessful trying the code suggested by fellow people (who claim it succeeeded) in the issue #13

[pongthang]

Hello,

I was unable to export finetuned (or any) model to ONNX.

I am using:

* pytorch 2.5.1

* torchaudio 2.5.1

* Python 3.10.10 (tried with 3.12.8 as well)

Code:

import torch

device = torch.device("cpu")

model = torch.hub.load(
    "IDRnD/ReDimNet",
    "ReDimNet",
    model_name="S",
    train_type="ft_mix",
    dataset="vb2+vox2+cnc",
).to(device)
model.eval()

input_sample = torch.randn(1, 32000).to(device)

dynamic_axes = {"feats": {0: "B"}, "embs": {0: "B"}}

torch.onnx.export(
    model,
    input_sample,
    "model_success.onnx",
    opset_version=17,
    input_names=["feats"],
    output_names=["embs"],
    dynamic_axes=dynamic_axes,
    verbose=True,
)

Received error:

Inputs:
        #0: 1060 defined in (%1060 : Float(*, *, *, strides=[16512, 16512, 1], requires_grad=0, device=cpu) = onnx::Pad[mode="reflect"](%1034, %1059), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram # /Users/aduomas/.pyenv/versions/3.10.10/lib/python3.10/site-packages/torch/nn/functional.py:5096:0
    )  (type 'Tensor')
        #1: 1070 defined in (%1070 : int[] = prim::ListConstruct(%1064, %1069), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram
    )  (type 'List[int]')
    Outputs:
        #0: 1071 defined in (%1071 : Float(*, *, strides=[16512, 1], requires_grad=0, device=cpu) = onnx::Reshape[allowzero=0](%1060, %1070), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram # /Users/aduomas/.pyenv/versions/3.10.10/lib/python3.10/site-packages/torch/functional.py:703:0
    )  (type 'Tensor')

Does the problem persist without dynamic_axes?

[pongthang]

One way you can try is to separate the preprocessing part - melspectrogram part. You can export the remaining networks. and implement the preprocessing part with your available library like librosa. Melspectrogram is the one that causes the error due its complex number.

[Aduomas]

Hello,
I was unable to export finetuned (or any) model to ONNX.
I am using:

* pytorch 2.5.1

* torchaudio 2.5.1

* Python 3.10.10 (tried with 3.12.8 as well)

Code:

import torch

device = torch.device("cpu")

model = torch.hub.load(
    "IDRnD/ReDimNet",
    "ReDimNet",
    model_name="S",
    train_type="ft_mix",
    dataset="vb2+vox2+cnc",
).to(device)
model.eval()

input_sample = torch.randn(1, 32000).to(device)

dynamic_axes = {"feats": {0: "B"}, "embs": {0: "B"}}

torch.onnx.export(
    model,
    input_sample,
    "model_success.onnx",
    opset_version=17,
    input_names=["feats"],
    output_names=["embs"],
    dynamic_axes=dynamic_axes,
    verbose=True,
)

Received error:

Inputs:
        #0: 1060 defined in (%1060 : Float(*, *, *, strides=[16512, 16512, 1], requires_grad=0, device=cpu) = onnx::Pad[mode="reflect"](%1034, %1059), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram # /Users/aduomas/.pyenv/versions/3.10.10/lib/python3.10/site-packages/torch/nn/functional.py:5096:0
    )  (type 'Tensor')
        #1: 1070 defined in (%1070 : int[] = prim::ListConstruct(%1064, %1069), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram
    )  (type 'List[int]')
    Outputs:
        #0: 1071 defined in (%1071 : Float(*, *, strides=[16512, 1], requires_grad=0, device=cpu) = onnx::Reshape[allowzero=0](%1060, %1070), scope: redimnet.model.ReDimNetWrap::/redimnet.layers.features.MelBanks::spec/torch.nn.modules.container.Sequential::torchfbank/torchaudio.transforms._transforms.MelSpectrogram::torchfbank.2/torchaudio.transforms._transforms.Spectrogram::spectrogram # /Users/aduomas/.pyenv/versions/3.10.10/lib/python3.10/site-packages/torch/functional.py:703:0
    )  (type 'Tensor')

Does the problem persist without dynamic_axes?

The problem persists without dyanamic_axes hyperparameter. Same error.

[Aduomas]

One way you can try is to separate the preprocessing part - melspectrogram part. You can export the remaining networks. and implement the preprocessing part with your available library like librosa. Melspectrogram is the one that causes the error due its complex number.

Is it a library version issue or has the redimNet preprocessing code has changed (since you were able to export it)

[knma1992]

I am also trying to export ReDimNet to onnx. Im curious if anyone is able to do it.

[Interspeaker2026]

Facing the same problem... are there any updates or workarounds?