Speed Up Re-compilation: A Handy Trick

Guides
,
1

If you’re frequently going through the conversion flow—whether due to continuous model re-training or re-quantization—you might find the final compilation step particularly time-consuming, especially when using the maximum performance mode. Fortunately, there’s a way to significantly reduce re-compilation time.

Here’s how:

  1. Save the compiled HAR file.
  2. Extract the contents of the HAR file. You can either extract it as any other tar file or use the hailo har extract command. The key file you need is MODEL_NAME.auto.alls.
  3. Use the extracted auto model script during re-compilation. If you are using the hailo CLI tool, you can pass the model script with the flag --model-script. With the Python API, re-load the model script before compiling.
  • Be aware that if the model has been re-trained, slight changes in allocation could occur. This may cause errors when using the entire auto.alls file.
  • To resolve this, start by removing all buffers(. commands from the file.
  • If issues persist, also remove the context_N.place( commands.