Support Knowledge Base

Accuracy Drops after Converting ONNX Model to Intermediate Representation (IR)

Content Type: Maintenance & Performance | Article ID: 000088868 | Last Reviewed: 07/07/2022

Description Resolution

Description

Converted ONNX model into IR.
Ran IR with Benchmark C++ Tool and its performance accuracy dropped 20% compared to running ONNX model with PyTorch.
Unable to determine how to pre-process image for better accuracy.

Resolution

Accuracy Checker supports set of preprocessors to process input data before model inference.

Change the type of supported preprocessors in configuration file to obtain the highest accuracy.

Related Products

This article applies to 2 products.

OpenVINO™ toolkit

Discontinued Products

Intel® Developer Cloud for the Edge

Need more help?

Contact support

Contact support