Time Required to Compile Quantized HAR Files to HEF Format?

General
1

Hello everyone,

I’m using Google Colab for my project, and I’ve encountered some challenges with the compilation time.

My first attempt to convert a quantized .har file to .hef format took over 11 hours, but I never received confirmation whether it was successful or not. The Google Colab runtime eventually restarted, so I lost track of the process. (See Image 1)
During my first attempt, the number of iterations reached around 270 (±10). After that, there was some workers, and the iteration process restarted.

Untitled
Untitled1920×1080 202 KB

For my second attempt, also on Google Colab, I’ve been monitoring the process more closely. Can anyone tell me approximately how many iterations this process should take to complete? How can I know if the compilation is close to finishing or not?

I’m also wondering if it would be better to compile on my local machine rather than using Google Colab? To be honest, I’m only using a small dataset of a few hundred images, so I’m concerned about whether compilation time would increase significantly with a larger dataset. Does dataset size influence compiling time ?
(See Image 2 for the iterations I’m referring to)

iteration
iteration722×718 42.4 KB

Any insights or recommendations would be greatly appreciated!

2

Hey @Vojin_Stevanovic,

I noticed your compilation issue and wanted to chime in. It looks like Google Colab timed out before your compilation finished. This happens quite often with HAR → HEF conversions since they’re CPU-intensive (you probably noticed the high CPU usage in Colab).

For these compilations, I’d really recommend either:

  • A powerful local machine
  • A cloud session without time limits

The compilation time varies based on several factors:

  • Your model’s complexity (depth, width, operations used)
  • Which Hailo device you’re targeting (Hailo-8, 8L, 15, 10H, etc.)
  • Compiler settings (optimization level, tiling, memory budget)

Dataset size does increase compile time, but it’s not actually the biggest factor.

As for iterations - typically you’ll see anywhere from 50-300 iterations depending on the model.

I would recommend running this overnight on a local workstation.