Emotion2Vec onnx?

[yairs-grandapp]

Hi all, is exporting the emotion2vec_plus models to onnx supported?
Is there a way to do this?

Thanks a lot!

[thewh1teagle]

Looking for onnx as well. I made some progress but this repository deleted the related issue where I wrote the notes :(

[LauraGPT]

Emotion2Vec ONNX export is available now. Support was added in #2359. I re-tested the current main (1d8080af) with the public iic/emotion2vec_plus_large checkpoint and ONNX Runtime CPU.

One current filename issue is easy to miss: the exporter writes a valid graph named emotion2vec without the .onnx suffix. Rename that file after export; no conversion is needed.

from pathlib import Path

import numpy as np
import onnx
import onnxruntime as ort
from funasr import AutoModel

output_dir = Path("./emotion2vec_onnx")
model = AutoModel(
    model="iic/emotion2vec_plus_large",
    hub="ms",
    device="cpu",
    disable_update=True,
)

export_dir = Path(
    model.export(
        type="onnx",
        quantize=False,
        device="cpu",
        output_dir=str(output_dir),
        opset_version=14,
    )
)

onnx_path = export_dir / "emotion2vec.onnx"
if not onnx_path.is_file():
    raw_path = export_dir / "emotion2vec"
    if not raw_path.is_file():
        raise FileNotFoundError(f"No Emotion2Vec ONNX graph found in {export_dir}")
    raw_path.rename(onnx_path)

graph = onnx.load(str(onnx_path), load_external_data=True)
onnx.checker.check_model(graph)

session = ort.InferenceSession(
    str(onnx_path),
    providers=["CPUExecutionProvider"],
)
waveform = np.random.randn(1, 16000).astype(np.float32)  # 16 kHz mono, [batch, samples]
features = session.run(["output"], {"input": waveform})[0]
print(features.shape)  # (1, 49, 1024)

Fresh verification results:

• exported graph: 648,972,628 bytes, opset 14;
• input/output names: input and output;
• 16,000 samples -> [1, 49, 1024], maximum absolute difference from PyTorch 6.20e-6;
• 32,000 samples -> [1, 99, 1024], maximum absolute difference 4.77e-6.

The exported interface is the dynamic-length Emotion2Vec feature extractor. Its output is frame-level 1024-dimensional features, not the final labels / scores dictionary returned by AutoModel.generate(). If you need the packaged emotion classifier and label post-processing, keep using the PyTorch AutoModel path for now.

The missing suffix is a naming bug in the current exporter, not a broken ONNX graph.