When changing an inference backend, record the runtime, driver, conversion tool, model file and execution configuration. Check where operations actually run and whether fallback occurs. Loading a model completes that step. Agreed samples are still needed to check results and resource use.
Compare outputs after numerical precision changes
Changing numerical precision can change numerical results and operator execution. ONNX Runtime quantization documentation states that quantization is not lossless and accuracy must be checked. Compare the original and deployed models using representative samples and task metrics. Save samples where the results differ, along with thresholds and evaluation versions.
Preprocessing defines the model input
Record color channel order, resizing, cropping, padding, normalization and tensor layout. Link original samples to actual input tensors and compare the transformed content. State whether preprocessing runs on the CPU or accelerator and whether copying is included in timing.
Verify optimization settings in the target environment
ONNX Runtime graph optimization documentation requires matching options and hardware when using offline optimization. Save the optimized model and configuration. Recheck function, time and resources in the target environment. Report model execution separately from the application path that includes video decoding, preprocessing and output processing.
See quantization documentation and graph optimization documentation.