Decreased Accuracy in Text Detection and Recognition Models after Upgrading to tritonclient 23.04-py3

[ashlinghosh]

Issue Summary:

After upgrading from tritonclient version 21.02-py3 to 23.04-py3, I noticed a decrease in accuracy for my text detection and recognition models, which are based on PaddleOCR. Prior to the upgrade, these models were delivering accurate results. The models are being used in their ONNX format for inference, not using the optimized plan version.

Steps to Reproduce:

Upgrade tritonclient to version 23.04-py3.
Use the text detection and recognition models based on PaddleOCR in their ONNX format.
Compare the results of text detection and recognition with those obtained using tritonclient version 21.02-py3.
Expected Behavior:

The text detection and recognition models should maintain or improve their accuracy after upgrading tritonclient.

Actual Behavior:

The accuracy of the text detection and recognition models has decreased significantly after the upgrade.

Additional Information:

The models are not in the optimized plan format; they are used directly in ONNX format.
Any specific configuration changes or updates in tritonclient version 23.04-py3 that could affect the accuracy of models based on PaddleOCR?
Are there any known compatibility issues or changes in behavior related to ONNX model inference in the newer tritonclient version?

Environment:

Tritonclient version: 23.04-py3
GPU : Tesla V100-PCIE-16GB