Troubleshooting

Not all models can be converted for the desired RVC Platform. We outline here some common issues that might arise during the conversion process.

Common Issues

Incorrect Conversion Parameters

Typically, mismatches occur in input/output tensor shapes, names, or types.SOLUTION: Thoroughly check your model and the meaning of all conversion parameters and adjust them accordingly. You can help yourself with the Netron tool to study the characteristics of an ONNX model.

Model Converts but Predicts Incorrectly

Conversion can succeed while runtime predictions are still wrong. This usually means the compiled model received the wrong input interpretation or that the outputs are being interpreted incorrectly.SOLUTION: Compare the same test image across the source framework, ONNX, and the converted model, then verify the following:

Input tensor name, shape, layout, and dtype match the source model.
encoding.from matches the source model's training-time color order and the runtime input matches encoding.to.
Normalization is applied exactly once. If conversion baked mean and scale into the exported model, do not repeat that normalization on the host.
An exported config.json with mean: null and scale: null can be correct; this often means preprocessing was already baked into the compiled model.
Resize, crop, and letterbox behavior still matches the original training or inference pipeline.
Class names, parser metadata, and heads configuration are present if the model requires semantic postprocessing.

Incompatible Model Formats

The source model format might not be supported by the conversion platform or the target format might not be compatible with certain features of the model.SOLUTION: It is advised to first convert the source model to ONNX format as it opens up the most options at conversion for the appropriate RVC Platform.

Unsupported Operations

The source model might contain operations that the conversion tool or the target model format do not support.SOLUTION: Consult the supported operations for the target platform you aim to run the conversion for. If the operation is not supported, consider replacing it with a supported alternative. If no alternative is available, consider breaking down the model into simpler components that can be converted separately and offload the problematic parts to on-host processing. You can help yourself with the onnx-modifier tool to introduce modifications to an ONNX model.

Model Architecture Issues

Complex or non-standard model architectures might cause conversion issues.SOLUTION: Simplify the model architecture if possible. Consider using the onnx.checker tool to identify potential issues in an ONNX model.

SNPE Compatibility

If you are troubleshooting RVC4 conversion or runtime issues, make sure the SNPE version used to export the model matches the SNPE runtime bundled with your Luxonis OS image and the DepthAI version you are running.Use the table below as a quick compatibility reference:

SNPE version	DepthAI version	Luxonis OS version
`2.32.6`	`3.0.0` up to `3.6.1`	`1.11.0` up to `1.25.3`
`2.41.0`	`3.6.1` and newer	`1.25.3` and newer

Models exported with an older SNPE version can generally run on newer SNPE runtimes. If the requested SNPE runtime is not available in your Luxonis OS image, DepthAI will automatically fall back to a lower runtime when possible. Models exported for a newer SNPE version than the device provides will not load; in that case, update Luxonis OS or re-export the model for a lower SNPE version.

ON THIS PAGE