Issues with custom Yolov8

[jeremyrbrown5]

Hello

We are training a yolov8 model to recognize rasberry pis. We are able to create the .tflite, test it using the predict function with 100% detection and high certainty, and push it to the Voxl 2 mini. However, no matter how we configure the voxl-tflite-server, we can't see it replicated in the voxl portal.

Some notes about our setup

For training we are using 75 images. 50 are for training, 15 are for valid, and 10 are for test

Our model-train.yaml is set to 30 epochs, 5 batchs, and 2 workers

our custom labels.txt for the tflite server is just 0 RasPi

we have used both the tracking_front and the Hires_misp_color cameras, with no such luck.

Inside of the deep learning documentation there is a section about the initialize_model_settings function in the main file. We cannot find this function on the voxl 2 mini. If the steps under 'Implementing your model in voxl-tflite-server' and 'Writing the model class' are critical to our success, we would really appreciate the function/file locations for the voxl 2 mini.

Thank you

[Zachary Lowell 0]

@jeremyrbrown5 I will plan on testing this out tomorrow - how are you validatign that it is detecting the rpis? Are you running voxl-inspect-detections tflite_data -a?

[jeremyrbrown5]

@Zachary-Lowell-0 we are using the yolo predict function inside the voxl-docker. I didn't know about voxl-inspect-detections, so I'll try that.

[Zachary Lowell 0]

@jeremyrbrown5 said in Issues with custom Yolov8:

@Zachary-Lowell-0 we are using the yolo predict function inside the voxl-docker. I didn't know about voxl-inspect-detections, so I'll try that.

Reply

If you are running your model directly on voxl-tflite-server then you can leverage the voxl sdk to detect any outputs from the model. That SDK is what is showing the image on voxl-portal. My guess is since this images arent showing on voxl-portal then you are having an issue during startup.

Can you run voxl-tflite-server directly on the command line and paste the output in here?

[Zachary Lowell 0]

@jeremyrbrown5 if you paste your model and upload it here i can download it and help troubleshoot the issue you are running into

[Zachary Lowell 0]

https://gitlab.com/voxl-public/support/voxl-train-yolov8

I am assuming you followed this instruction set for training your model?

[jeremyrbrown5]

@Zachary-Lowell-0

stinger (D0013):/$ voxl-tflite0-server
bash: voxl-tflite0-server: command not found
stinger (D0013):/$ voxl-tflite-server

skip_n_frames: 5

model: /usr/bin/dnn/best_float16.tflite

input_pipe: /run/mpa/hires_misp_color/

delegate: gpu

allow_multiple: false

output_pipe_prefix: mobilenet

existing instance of voxl-tflite-server found, attempting to stop it
WARNING: Unknown model type provided! Defaulting post-process to object detection.
INFO: Created TensorFlow Lite delegate for GPU.

[jeremyrbrown5]

@Zachary-Lowell-0 Yes I have been using the guides linked on the modalai website

[jeremyrbrown5]

@jeremyrbrown5 here is a link to our current model .tflite file

[jeremyrbrown5]

@jeremyrbrown5 we figured it out, we just changed the name of the model to yolov8n_float16.tflite. Pretty spot on now!