[Data compile] Which layer has the most calculations?

General
, ,
1

I am trying to adopt SR model to hailo model.
And there are some issues during the data compile.

During optimization and compilation, a lot of GPU memory is used or mapping failure occurs.
I would like to know which layer requires the most calculations.

Also, thank you for answering every inquiry.

2

Hey @hwan136 ,

1. Profile Your Model

To identify the layers with high computational costs or memory requirements in your ONNX model, use the Hailo Dataflow Compiler’s profiling feature:

hailo profiler --input your_model.onnx --output profile_report.json

The generated profile_report.json file will provide detailed information about each layer, including the FLOPS (floating-point operations) and memory usage. Review this report to pinpoint the most resource-intensive layers in your model.

2. Identify Computational Bottlenecks

When analyzing the profile report, pay special attention to the following types of layers, as they are typically the most computationally demanding:

  • Convolutional layers, especially those with large kernel sizes or strides.

    • Look for layers with high input/output tensor sizes.
    • Check for layers with large kernel sizes or dilations.

  • Upsampling layers or pixel-shuffle operations in super-resolution (SR) models.

    • These layers often introduce significant memory overhead.

For ready super-resolution models, please refer to the Hailo Model Zoo documentation on SR models: HAILO8_super_resolution.rst

If you encounter any errors or issues during the optimization process, please provide the specific error messages or logs you receive. This will help us better understand the problem and provide more targeted assistance.

Best regards,
Omria