Hi,
I’m converting a custom YOLO26 model to .hef using Hailo Model Zoo on Google Colab (Tesla T4). The process completes ONNX translation, mixed precision, and statistics collection successfully, but gets stuck at the Adaround step.
Log excerpt:
[info] The algorithm Adaround will use up to 36.91 GB of storage space
[info] Using dataset with 1024 entries for Adaround
[warning] DALI is not installed, using tensorflow dataset for layer by layer train.
[warning] Dataset isn't shuffled without DALI.
Then it hangs indefinitely. I have to interrupt it with Ctrl+C.
My questions:
- Is 37 GB storage requirement normal for Adaround on YOLO26?
- Does missing DALI cause it to hang, or just slow down?
- Can I skip Adaround without major accuracy loss?
- Any recommended workarounds for Colab’s disk/memory limits?
Thanks!
Adaround is quite heavy on memory, you can lower the batch size in the modelscript (32 is default). Installing DALI gave me some performance, but don’t expect crazy speedups. You can try finetune instead of adaround, this gives similar accuracy for my model (NanoDet).
I did a test conversion of yolo26n on my machine using hailomz from the AI Software Suite Docker 2026-07, and it worked as expected.
The optimization should still work without DALI. During my test run, it progressed normally, although it was very slow. I stopped the process after about an hour because of the runtime.
As @josef.gugglberger mentioned, the speedup from DALI is not dramatic, but it was enough to reduce the conversion time of yolo26n to about 2.5 hours on my machine with an RTX 4090.
It’s definitely worth trying. If the resulting accuracy degradation is acceptable for your application, you can skip the Adaround step and save a significant amount of conversion time.
Unfortunately, model conversion and AI model training are resource-intensive. A powerful machine with plenty of RAM and a capable GPU is highly recommended for a smooth conversion process.