skip to content
arpit agrawal
note1 min

What 100 TensorRT conversions taught me about ONNX

Most failures happened before TensorRT was involved.

Most conversion failures happened before TensorRT was involved at all: dynamic shapes that were never declared, custom ops with no ONNX equivalent, and exports that only worked for the batch size they were traced with.

Once the export was clean, quantization was mostly a question of calibration data.