Below are some of the troubleshooting guidelines for OSRT model compilation and inference. This section is common across the all the OSRT (TFlite / ONNX runtime /TVM-DRL )

Set debug_level > 0 to enable verbose debug log during import and inference
The import process generates visualization of the model in artifacts folder. This can be used to understand the number sub graphs offloaded to C7x-MMA, layers in each subgraph and layers processed on ARM processor
If there is any issues observed during model compilation process, enable ARM only mode by passing “-d” as argument to default model compilation script provided Example
python3 onnxrt_ep.py -d
If the model works fine with ARM only mode but fails to compile with C7x-MMA offload, try dispatching some of the layers (less commonly used layer type) to ARM by using “deny_list” option
If accuracy of the model is observed lower than expected, try to import the model with 16-bit inference. This can be enabled by setting tensor_bits=16. If the 16-bit mode works fine, then fine tune the advanced_options:calibration to work with 8-bit mode. If the 16-bit mode also fails to meet accuracy requirement, then try with “deny_list” option as described in previous step.