YoloExtended Parser OAK 4 D Pro Error

KevinSani

Hello there @jakaskerl, @jakob I am trying to use the YoloExtended Parser to parse through my custom model for object detection, the nn archive I created was able to convert the model properly but when I try to parse thorugh the model I am getting this particular error
(venv):~/depthai-core/examples/python/DetectionNetwork$ python3 yolo_test.py

terminate called after throwing an instance of 'pybind11::error_already_set'

what(): RuntimeError: Tensor does not exist

At:

/home/user/depthai-core/venv/lib/python3.10/site-packages/depthai_nodes/node/parsers/yolo.py(289): <listcomp>

/home/user/depthai-core/venv/lib/python3.10/site-packages/depthai_nodes/node/parsers/yolo.py(288): run

I am pretty sure the way I defined my archive config file maybe wrong can someone please give me the proper way to design the archive config file for a YoloExtended Parser or atleast point me in the right direction to go

These are my model details

main_graph

Inputs

tensor: float32[1,3,896,896]

Outputs

tensor: float32[1,10,16464]

Metrics

3089157

KlemenSkrlj

Hi @KevinSani ,
based on the model information you shared it looks like your output shapes are incorrect. The thing is that the YoloExtendedParser expects a model that has 3 outputs where some of the operations right at the end of the original model are prunned. If you trained your model using e.g. Ultralytics or YOLOv6 official website (or some other well known framework) then we should already support the conversion process going from pytorch weights to format that can be ran on the device. To try this out please check out the HubAI where you'll see a button for "Convert model - YOLO". This should then generated you NNArchive together with correctly exported model that you can then use out of the box with YoloExtendedParser.
Let me know if that works,
Klemen

KevinSani

Hello KlemenSkrlj
When I tried converting my custom model I got this error on hubai conversion logs

Conversion process logs

[2025-06-11 14:53:31.105934+00:00] {pod_manager.py:472} INFO - [base] INFO Loading model... main.py:118 [2025-06-11 14:53:31.106512+00:00] {pod_manager.py:472} INFO - [base] mkdir -p failed for path /home/non-root/.config/matplotlib: [Errno 13] Permission denied: '/home/non-root' [2025-06-11 14:53:33.228594+00:00] {pod_manager.py:472} INFO - [base] Matplotlib created a temporary cache directory at /tmp/matplotlib-lc00oc74 because there was an issue with the default path (/home/non-root/.config/matplotlib); it is highly recommended to set the MPLCONFIGDIR environment variable to a writable directory, in particular to speed up the import of Matplotlib and to better support multiprocessing. [2025-06-11 14:53:33.229091+00:00] {pod_manager.py:472} INFO - [base] WARNING /usr/local/lib/python3.11/site-packages/tools/yolo/yolov5 logging.py:92 [2025-06-11 14:53:33.229664+00:00] {pod_manager.py:472} INFO - [base] `torch.load` with `weights_only=False` (the current [2025-06-11 14:53:33.230152+00:00] {pod_manager.py:472} INFO - [base] default value), which uses the default pickle module [2025-06-11 14:53:33.230344+00:00] {pod_manager.py:472} INFO - [base] implicitly. It is possible to construct malicious pickle [2025-06-11 14:53:33.230628+00:00] {pod_manager.py:472} INFO - [base] data which will execute arbitrary code during unpickling [2025-06-11 14:53:33.230937+00:00] {pod_manager.py:472} INFO - [base] (See [2025-06-11 14:53:33.231262+00:00] {pod_manager.py:472} INFO - [base] https://github.com/pytorch/pytorch/blob/main/SECURITY.md# [2025-06-11 14:53:33.232426+00:00] {pod_manager.py:472} INFO - [base] untrusted-models for more details). In a future release, [2025-06-11 14:53:33.232952+00:00] {pod_manager.py:472} INFO - [base] the default value for `weights_only` will be flipped to [2025-06-11 14:53:33.233365+00:00] {pod_manager.py:472} INFO - [base] `True`. This limits the functions that could be executed [2025-06-11 14:53:33.233676+00:00] {pod_manager.py:472} INFO - [base] during unpickling. Arbitrary objects will no longer be [2025-06-11 14:53:33.233978+00:00] {pod_manager.py:472} INFO - [base] allowed to be loaded via this mode unless they are [2025-06-11 14:53:33.234554+00:00] {pod_manager.py:472} INFO - [base] explicitly allowlisted by the user via [2025-06-11 14:53:33.235051+00:00] {pod_manager.py:472} INFO - [base] `torch.serialization.add_safe_globals`. We recommend you [2025-06-11 14:53:33.235063+00:00] {pod_manager.py:472} INFO - [base] start setting `weights_only=True` for any use case where [2025-06-11 14:53:33.236173+00:00] {pod_manager.py:472} INFO - [base] you don't have full control of the loaded file. Please [2025-06-11 14:53:33.236532+00:00] {pod_manager.py:472} INFO - [base] open an issue on GitHub for any issues related to this [2025-06-11 14:53:33.236864+00:00] {pod_manager.py:472} INFO - [base] experimental feature. [2025-06-11 14:53:33.237287+00:00] {pod_manager.py:472} INFO - [base] ckpt = torch.load(attempt_download(w), [2025-06-11 14:53:33.238045+00:00] {pod_manager.py:472} INFO - [base] map_location='cpu') # load [2025-06-11 14:53:35.339770+00:00] {pod_manager.py:490} INFO - [base] ERROR Error creating exporter: No module named 'ultralytics' main.py:156

**But when I uploaded a onnx model it properly converted but when i used that in the program I got this error of no heads defined in the nnarchive

depthai-core/examples/python/DetectionNetwork$ python3 yolo_test.py**

Traceback (most recent call last):

File "/home/depthai-core/examples/python/DetectionNetwork/yolo_test.py", line 17, in <module>

**detection_nn = pipeline.create(ParsingNeuralNetwork).build(**

File "/home/depthai-core/venv/lib/python3.10/site-packages/depthai_nodes/node/parsing_neural_network.py", line 278, in build

**self._updateParsers(self._nn_archive)**

File "/home/depthai-core/venv/lib/python3.10/site-packages/depthai_nodes/node/parsing_neural_network.py", line 311, in _updateParsers

**self._parsers = self._getParserNodes(nnArchive)**

File "/home/depthai-core/venv/lib/python3.10/site-packages/depthai_nodes/node/parsing_neural_network.py", line 321, in _getParserNodes

**parsers = self._generateParsers(parser_generator, nnArchive)**

File "/home/depthai-core/venv/lib/python3.10/site-packages/depthai_nodes/node/parsing_neural_network.py", line 332, in _generateParsers

**return parserGenerator.build(nnArchive)**

File "/home/depthai-core/venv/lib/python3.10/site-packages/depthai_nodes/node/parser_generator.py", line 42, in build

**raise ValueError("No heads defined in the NN Archive.")**

**ValueError: No heads defined in the NN Archive.

I wanted to remind you that I am converting for an RVC4 device an OAK 4 D PRO camera**

KlemenSkrlj

I see. Based on the error I assume you have trained the model with Ultralytics YoloV5 repo? Would you be able to share the .pt file with us (can be privately over gmail as well at klemen.skrlj@luxonis.com ) or tell us which configuration you've used (nano, small, etc.) so we can use it for debugging on our side?

KevinSani

Its actually a yolov8n model, if possible what is the best parser to use this model for this type of output

KlemenSkrlj

Hm I just tested converting yolov8n model from official Ultralytics website here and the conversion worked fine. Could you share some more details of where and how did you train this model?

KevinSani

KlemenSkrlj
I have sent the model info and the model to your email.

KevinSani

Hey @KlemenSkrlj
I used the converted model you just sent but I am getting this error

[1443462065] [192.168.0.152] [1749743773.947] [NeuralNetwork(2)] [error] Couldn't build the model.

[2025-06-12 10:56:14.818] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_2_passthrough' (X_LINK_ERROR)'

[2025-06-12 10:56:14.818] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_4_out' (X_LINK_ERROR)'

[2025-06-12 10:56:17.867] [warning] FW crashed - trying to get out the crash dump

[2025-06-12 10:56:20.867] [warning] Getting the crash dump out - this can take up to a minute, because it first needs to be compressed.

[2025-06-12 10:56:23.882] [depthai] [error] Device with id 1443462065 has crashed. Crash dump logs are stored in: /home/depthai-core/examples/python/DetectionNetwork/.cache/depthai/crashdumps/084a3702eb64eb6504036e79d44a5ca5948847df/core_dump_0.0.1+c21e47f41f34e3d818fb11cfb44726c5edaf2695_20250612_105621.tar.gz - please report to developers.

[1443462065] [192.168.0.152] [1749743793.834] [host] [warning] Closed connection

[1443462065] [192.168.0.152] [1749743793.834] [host] [warning] Attempting to reconnect. Timeout is 10000ms

[1443462065] [192.168.0.152] [1749743794.815] [NeuralNetwork(2)] [error] Couldn't build the model.

[1443462065] [192.168.0.152] [1749743794.842] [host] [warning] Reconnection successful

[2025-06-12 10:56:35.738] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_4_out' (X_LINK_ERROR)'

[2025-06-12 10:56:35.738] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_2_passthrough' (X_LINK_ERROR)'

[2025-06-12 10:56:39.863] [warning] FW crashed - trying to get out the crash dump

[2025-06-12 10:56:42.863] [warning] Getting the crash dump out - this can take up to a minute, because it first needs to be compressed.

[2025-06-12 10:56:45.944] [depthai] [error] Device with id 1443462065 has crashed. Crash dump logs are stored in: /home/depthai-core/examples/python/DetectionNetwork/.cache/depthai/crashdumps/e4d87a18b7a24c11de9f1b5a30d0013e295a2c75/core_dump_0.0.1+c21e47f41f34e3d818fb11cfb44726c5edaf2695_20250612_105643.tar.gz - please report to developers.

[1443462065] [192.168.0.152] [1749743817.167] [host] [warning] Closed connection

is this a programming error ???

this is where it breaks
detection_nn = pipeline.create(ParsingNeuralNetwork).build(

camera_node.requestOutput((896, 896), type=img_frame_type, fps=30),

nn_archive

)

I used the same programming structure as in this repo https://github.com/luxonis/depthai-ml-training/blob/main/conversion/pytorch_conversion.ipynb

KlemenSkrlj

Based on this log:
[1443462065] [192.168.0.152] [1749743773.947] [NeuralNetwork(2)] [error] Couldn't build the model.
It seems like you might not have build the model for RVC4? To quickly double check this the .tar.gz file you got and check what is inside it? There should be: buildinfo.json, config.json and a .dlc file which is the actual model exported for RVC4.

KevinSani

Hey @KlemenSkrlj So can you tell me the the model path that should be specified using model_path variable. I actually used the model you have converted and sent.

KlemenSkrlj

The model_path must be a path to the .tar.gz file. If you are using the files I sent in you in the MRE then the model_path would be ./mre/last.rvc4.tar.xz.
Just an extra note: The tutorial you've took as reference is for classification model where as you have an object detection model. The pipeline itself doesn't differ, there will be just some modifications needed in the very last part, when you want to visualize the model (the visualize_classification function doesn't apply to you). You can check out the last part of the detection_model tutorial or some of the oak-examples (here is the generic example which you can use or more advanced object-detection type of examples).

KevinSani

Hello @KlemenSkrlj
MODEL_PATH = "mre/last.rvc4.tar.xz"

device = dai.Device()

img_frame_type = dai.ImgFrame.Type.BGR888i if device.getPlatform().name == "RVC4" else dai.ImgFrame.Type.BGR888p

with dai.Pipeline(device) as pipeline:

camera_node = pipeline.create(dai.node.Camera).build()

nn_archive = dai.NNArchive(MODEL_PATH)

detection_nn = pipeline.create(ParsingNeuralNetwork).build(

camera_node.requestOutput((896, 896), type=img_frame_type, fps=30),

nn_archive

)

is this correct this is how i am trying to run the program and I get the above error which i specified. The code breaks at
this is where it breaks
detection_nn = pipeline.create(ParsingNeuralNetwork).build(

camera_node.requestOutput((896, 896), type=img_frame_type, fps=30),

nn_archive

)

Did you test the model in code to see if its working??

KlemenSkrlj

You are not being careful about the input shape. Here you are requesting (896,896) input shape from the camere but the input shape of the models I exported for you is 640x640. This information you can see from the config.json inside the .tar.gz file. If you wish you can export for other input shapes as well using Hub conversion mentioned previously.

And yes, if you fix this, the the snippet you shared works and was tested to work with your model. To avoid the issue of having to manually set the input shape in the DepthAI script I'm sharing a more easy to use snippet below (please install depthai==3.0.0rc2 and depthai-nodes==0.3.0 to use it). The only thing to change is the IP to set it to yours device IP.

import depthai as dai
from depthai_nodes.node import ParsingNeuralNetwork
MODEL_PATH = "mre/last.rvc4.tar.xz"
device = dai.Device(dai.DeviceInfo("x.x.x.x")) # TODO: Change to IP of your RVC4 device
with dai.Pipeline(device) as pipeline:
    camera_node = pipeline.create(dai.node.Camera).build()
    nn_archive = dai.NNArchive(MODEL_PATH)
    detection_nn = pipeline.create(ParsingNeuralNetwork).build(
    camera_node,nn_archive)
    parser_output_queue = detection_nn.out.createOutputQueue()
    pipeline.start()
    while pipeline.isRunning():
        parser_msg: dai.ImgDetections = parser_output_queue.get()
        print(parser_msg.detections)

KevinSani

Hello @jakaskerl , @KlemenSkrlj thanks a lot for your help, i am currently using a new model which has 2 outputs instead of a single one and the same code used here doesnt seem to work should i use the detection parser to run this new model???

KlemenSkrlj

If your model doesn't conform to what is expected by the parser (in this case YOLOExtendedParser) then you should probably build your own custom HostNode where you'll have full control over how the postprocessiong is done. In that case I advise you to use dai.node.NeuralNetwork directly and link its output to your custom parser node. Feel free to check out the DepthAI HostNode documentation or even the depthai-nodes source code to help you kickstart the development od your parser.

KevinSani

I am sorry @KlemenSkrlj I am a bit confused can you explain a bit more clearly this is my model configuration, any information you share would be really helpful
-

tensor: float32[1,3,896,896]

Outputs

tensor: float32[1,39,16464]

tensor: **float32[1,32,224,224]
when I ran the program above with my new model this is the output i am getting this one doesnt have detections is this a model issue
**
while pipeline.isRunning():

parser_msg: dai.ImgDetections = parser_output_queue.get()

print(parser_msg)**

sequenceNum: 16**

ts:

**`nsec: 804520000`**

**`sec: 12813`**

tsDevice:

**`nsec: 0`**

**`sec: 0`**

sequenceNum: 17

ts:

**`nsec: 837854000`**

**`sec: 12813`**

tsDevice:

**`nsec: 0`**

**`sec: 0`**

thanks for your help

KlemenSkrlj

I would need more context into what this new model is and what kind of outputs does it have ie. Is it object detection? What is the second head? What is the semantic meaning? etc.
But generally I assume you have some code that transforms this outputs to something meaningful (like bounding boxes for object detection). You can convert the model as is using Hub modelconversion and then when developing DepthAI pipeline you can put all the postprocessing code in a HostNode. The pipeline will than match the one you have currently where you have some script that runs inputs through the model and then you have some postprocessing function that turns raw model outputs to a more meaningful form.

KevinSani

Hello @KlemenSkrlj just fixed the problem thanks for your help your documentation with codes were really helpful

KevinSani

Hey @KlemenSkrlj, @jakaskerl I am using an oak4d pro camera and iam rotating the frame 90 degrees clockwise I am using this code from this repo luxonis/depthai-ml-trainingblob/main/training/train_instance_segmentation_model.ipynb
when I just run the code normally its working but when i rotate the frame the output seems to be completely wrong I tried rotating the mask and the bounding boxes too but that doesnt seem to fix the issue can you please help me or point me towards the right resources to solve this problem

KlemenSkrlj

Not sure if I understand correctly what you are trying to do. But if you want to predict on rotated images then you'll have to add an ImageManip node between the camera and the neural network nodes which will do the rotation (check here how to set the ImageManip config). With a pipeline like this (camere->image manip->NN) then the NeuralNetwork will predict on rotated frame and the .passthrough output of the NN will also be rotated frame so you can use it to visualize the predictions on it.