YOLOv5s

[m7md5303]

Is your flow valid only for face recognition? In other words, I want to deploy YOLOv5s on ZCU102 but as you know it is for object detection not face recognition. Can I train your network for object detection or should I edit YOLOv5 before the deployment?
Thanks in adavnce

[sefaburakokcu]

Hi @m7md5303,

Thank you for your question!

The repository is indeed for object detection, and we specifically focused on a single-class object detection task: face detection. However, you can modify the number of classes and use your own dataset to train the model for your specific object detection task.

That said, please note that we used a simplified YOLO model to ensure compatibility with FINN for deployment purposes. As a result, the YOLOv5 model architecture might not be directly compatible without additional modifications. You may need to adjust the model architecture or adapt it to work with FINN if you plan to deploy YOLOv5 on the ZCU102.

[m7md5303]

@sefaburakokcu Thank you very much for this clarifying answer. I will then try using your network after these small modifications.
Best Regards

[lilheng]

Hi @sefaburakokcu, I am currently trying to quantize a tiny yolo model 3 with generic object detection that will be deployed onto FINN and eventually on Kria FPGA board. I have trained the model using this repository, but I am not sure how does the quantization of the model occur.
Are there any help you can offer ?

[m7md5303]

@sefaburakokcu would you please provide the code for the LPYOLO network for retratining on custom datset. I would appreciate it if you made it available.
Thanks in advance

[sefaburakokcu]

Hi @m7md5303,

You can run

python train.py --data your_dataset.yaml --cfg models/hub/yolov1-tiny-quant.yaml --weights '' --batch-size 128

for training LPYOLO on your custom dataset. You should also update number of classes in yolov1-tiny-quant.yaml according to your dataset.

[m7md5303]

Many thanks, Mr. @sefaburakokcu