How can the recognition model export in onnx format support variable length? The training size is 32 * 128, and using onnx model to predict other size formats may cause errors, such as 32 * 256 or 32 * X, where position encoding has been disabled

[dlml]

No description provided.

[Topdu]

We are working on exporting the onnx model and testing it, stay tuned.