HubAI-compiled YOLO model missing detection heads and others

NilStohr

Hi Forum,

I have been struggling with this for three days now, and I would really appreciate some help.

I have an OAK-D Pro camera. The camera is installed, and I have also installed the DepthAI API 3.7.1 for Python. Everything works fine with the default examples.

I have been running this example:

"https://docs.luxonis.com/software-v3/depthai/examples/spatial_detection_network/spatial_detection/"

It works out of the box. I also modified the code to use dai.NNArchive instead of dai.NNModelDescription.

I downloaded the model:

yolov6n-r2-288x512.tar.xz

from the Luxonis Zoo:

"https://models.luxonis.com/luxonis/yolov6-nano/aim_RFFiRGVUcFAVE894kxRbRm?backTo=%2F"

I did this just to try fetching the model myself. The yolov6n-r2-288x512.tar.xz archive contains a .superblob file and a config.json file. Everything works fine with this precompiled model, so that part is all good.

Now to the part where I have been pulling my hair out.

My goal is to download, and later train, a non-compiled YOLOv6 Nano model. I thought that, as a starting point, it would be a good idea to download a non-compiled model, compile it myself, and then use it with the example above.

I downloaded and compiled an Ultralytics YOLOv6 Nano model using Python, Ultralytics, and HubAI, as shown below:

$$
from ultralytics import YOLO
from hubai_sdk import HubAIClient
import os

model_name = "yolov6n"
model = YOLO(f"{model_name}.yaml")

print("Loading and exporting " + model_name)
onnx_path = model.export(format="onnx", opset=11, imgsz=320)

key = os.getenv("HUBAI_API_KEY")
client = HubAIClient(api_key=key)

response = client.convert.RVC2(
path=onnx_path,
name=model_name,
is_yolo=True,
superblob=True,
number_of_shaves=5,
input_shape=[320, 320],
)
$$

The model compiles successfully, and I get a file called:

yolov6n.rvc2.tar.xz

I then tried to run the same example again:

"https://docs.luxonis.com/software-v3/depthai/examples/spatial_detection_network/spatial_detection/"

This worked with the Luxonis Zoo model, yolov6n-r2-288x512.tar.xz, but when I run it with the model I manually compiled, yolov6n.rvc2.tar.xz, I get this error:

$$
RuntimeError: NNArchive should contain at least one detection head (YOLO or Mobilenet-SSD).
$$

Looking closer at the config.json file inside yolov6n.rvc2.tar.xz, I can see that "heads": [] is empty. I also checked the config.json file inside yolov6n-r2-288x512.tar.xz, and in that file the heads section is not empty. So the two config.json files are clearly not the same.

My questions are:

Why does HubAI generate a config.json file with missing content?

Am I doing the whole flow completely wrong?

Overall, what I want to do is:

Use a basic YOLOv6 Nano model.
Train it to recognize shoes using the Ultralytics Python API.
Compile it using HubAI.
Run it live using the DepthAI spatial detection example.

I also tried fetching a YOLO11 Nano model and training it with a YOLO dataset using the Ultralytics Python API. I validated that the model worked using model.predict. I then tried to compile the model using HubAI, but it also produced a config.json file with missing content in the heads section.

So in general, I do not really understand why HubAI compiles the models with missing metadata.

I hope I am not supposed to manually expand these JSON files. I have been looking into doing that, but I cannot find proper documentation, and things keep failing when I try.

I would really appreciate any help, as I am close to giving up.

Thanks in advance.

Oskar-Luxonis

cc @KlemenSkrlj

KlemenSkrlj

Hi @NilStohr ,
The main problematic in this flow is how you are converting the model to get the RVC2 exported version. Since this is a YOLO model you want to follow the YOLO export path. This is described in our "Running YOLO Neural Networks Onboard OAK4" tutorial for HubAI interactive flow.
But since you are doing it through the HubAI SDK the main point is that you want to take the .pt file as input, not the ONNX. This is important because then we know how to parse the .pt file into a valid ONNX NNArchive with heads section specified as well and we can carry this automatically over to RVC2 exported NNArrchive.

So in general, I do not really understand why HubAI compiles the models with missing metadata.

If you do conversion of YOLO through the expected YOLO flow with .pt file as a starting point (not ONNX) then we can get from the .pt file the metadata needed to create the proper JSON file with heads specified. In the case when input is an ONNX though (as is in your current case) then we can't know automatically what kind of model this is, what kind of parser would it use, etc. So there is essentially no way for us to know what one would want defined in the heads and in that case one would need to specify it manually. But this is not your case, so no worries. Using .pt file directly should solve all your issues.

I hope I am not supposed to manually expand these JSON files. I have been looking into doing that, but I cannot find proper documentation, and things keep failing when I try.

Not that you need it for your case but just as extra context: Here is the NNArchive documentation which would be what you would need if you would need to define the JSON file manually.

Let me know if this makes more sense now and if you have any more issues/questions.
Best,
Klemen

NilStohr

Thanks a lot for the response.

Providing the HubAI model as a .pt file produced a config.json with the heads content, so that part is good. I modified my Python code as shown below:

$$
model_name = "yolov6n"
pt_path = os.path.join(WS_PATH, f"{model_name}.pt")
print("Converting " + model_name + " → RVC2 blob via HubAI")

key = os.getenv("HUBAI_API_KEY")
client = HubAIClient(api_key=key)

response = client.convert.RVC2(
path=pt_path,
name=model_name,
is_yolo=True,
yolo_version="yolov8",
superblob=True,
number_of_shaves=5, # 5 CMX slices: matches RGBD Spatial pipeline allocation
input_shape=[320, 320], # must match training resolution
)
$$

After that, I tried the example again:

https://docs.luxonis.com/software-v3/depthai/examples/spatial_detection_network/spatial_detection/

I made only a small modification to use dai.NNArchive. Please see the attached file

modified-spatial-detection-network.bin

#!/usr/bin/env python3
import argparse
from pathlib import Path
import cv2
import depthai as dai

Error previewing file. It may have been deleted, or the provided file ID is invalid.

When running the example with the model compiled by HubAI "yolov6n.rvc2.tar.xz", with the correct metadata in the config, the program almost crashes, lagging extremely (no id boxes shown either).

[2026-06-25 14:13:44.670] [ThreadedNode] [warning] Sync node has been trying to sync for 51 messages, but the messages are still not in sync.
[2026-06-25 14:13:44.670] [ThreadedNode] [warning] Output depthPreview timestamp is 16672267 ms

[2026-06-25 14:13:44.670] [ThreadedNode] [warning] Output detections timestamp is 16672217 ms

[2026-06-25 14:13:44.670] [ThreadedNode] [warning] Output rgbPreview timestamp is 16672217 ms

[2026-06-25 14:13:44.670] [ThreadedNode] [warning] Sync node has been trying to sync for 52 messages, but the messages are still not in sync.

However, if I run the same example with the precompiled model downloaded from the zoo, YOLOv6_Nano-R2_COCO_512x288.rvc2.tar.xz, it does not crash, and works perfectly. What is up here?

So, the missing content from config.json appears to be fixed by providing the .pt file instead of the ONNX file, but now the program crashes. I suspect model to be the problem.

Could you please inspect or review the attached Python script? It only differs slightly from the official example, mainly in that it uses an archive.

Thanks very much in advance.

KlemenSkrlj

Hey @NilStohr ,
Just to confirm:

You're using the nano variant of YOLOv6?
And you are using stereo as the depthSource?
What kind of RVC2 device do you have?

In the meantime I'll try to reproduce you issue but answering these questions will provide more context around it.
Best,
Klemen