trouble with deployment of custom YOLO ONNX model

seems like it’s just a warning but I don’t know if this is the reason of my error

HI ​@tchretien!

As you say, Voyager defaults to batch size 1 unless explicitly configured otherwise in the model YAML. So I don’t think your YOLO model with batch size 8 will work. But if you can re-export it with batch_size=1 you could be onto a winner.

There's no built-in pipeline functionality for batching multiple images before inference. If you need higher throughput, you can use the SDK's multi-core capabilities at the application level, but that's separate from model batch size.

For deploying your custom ONNX model, maybe check out:

• The Custom Model Tutorial
• The t1-simplest-onnx tutorial for basic ONNX examples

Let me know how it goes!

To add to Spanner’s note: batch size is something set at compile time. We support batch sizes up to 4 since our Metis AIPU has four cores. Here’s an overview of how to configure this, look for the “Multi-core modes” section.

If your model input shape is size 1, then our SDK will create four instances of the model, one per core, and handle frame batching to each model.

Thanks,

David