nadaved1
commented
1 year ago Hi @MyraBaba , There are no re-trainging dockers for yolov7 now, but as you raise this request we will see if/when we can these. As you you've mentioned that you've already trained and exported to ONNX your own model, you can try to compile it to HEF (network binary format that can run on a Hailo-8 device) using the model-zoo recipe for YOLOv7, or just the DFC.
We only have this specific model for vehicle detection, but as you've mentioned it's possible to use a re-trainging docker with visdrone to achieve a new detector.
Tappas is end-2-end application including the image acquisition, pre/post processing, visual etc. Some networks could fit into existing pipelines (e.g. detection), for other the pipeline would need to be further developed, like implementing a post processing function. Again, if there is some specific network/pipeline that is of interest to you please let us know, and we will consider adding it.
I encourage you to reach out to us through the official support system through this link for a more thorough support. hailo.ai/contact-customer-support/
It's not always easy to recreate the same reported numbers that a Git or a paper reports.
Hi,
We would possibly test Hailo-8 for our upcoming projects and need to understand the further:
Is there any retraining for hailo-8 yolov7 docker and configs? we have trained yolov7 models (tiny , v7 , w , e ) for visdrone which we are happily used . Both pt and onnx are available . Can we convert those model to hailo-8 ? is there any restrictions ?
I saw yolov7 model in the model_zoo . Is there any benchmark for fram per second for those ?
Hailo-8 trained "vehicle detection" looks very good. but only one class.. Is there any vehicle detection like visdrone class?
There is "Networks available in TAPPAS are marked with star" models. Are other models not available for use in Tappas ? what does it mean ?
PS: Official yolov7 accuracy and yours mAP is different (no quantized version) https://github.com/WongKinYiu/yolov7
Best