Darknet: Is it possible to train a YOLO model using Mobilenet backend ?

Created on 21 Nov 2017 · 1Comment · Source: AlexeyAB/darknet

Hi @AlexeyAB
I found out some articles that implement yolo v2 using Mobilenet as a backend.
ex: https://github.com/experiencor/basic-yolo-keras
However this is using keras so I'm wondering if it's possible to do the same thing using darknet ?
Thanks

Source

oishi89

Most helpful comment

@oishi89 Hi,

MobileNet requires Depth-wise convolutional layers: https://arxiv.org/abs/1704.04861v1
So, if you can added Depth-wise to the Darknet (or get it from there): https://github.com/AlexeyAB/darknet/issues/112
Then you can change several first layers in yolo-voc.2.0.cfg to layers from MobileNet, in such way as I did it for DenseNet and MSRA(ResNet): https://github.com/AlexeyAB/darknet/issues/179#issuecomment-330047738
But there isn't MobileNet wegihts for classification network, so you will train your own detection network without pre-trained weights: ./darknet detector train obj.data mobile-yolo.cfg

AlexeyAB on 21 Nov 2017

❤3 🚀1

>All comments

@oishi89 Hi,

MobileNet requires Depth-wise convolutional layers: https://arxiv.org/abs/1704.04861v1
So, if you can added Depth-wise to the Darknet (or get it from there): https://github.com/AlexeyAB/darknet/issues/112
Then you can change several first layers in yolo-voc.2.0.cfg to layers from MobileNet, in such way as I did it for DenseNet and MSRA(ResNet): https://github.com/AlexeyAB/darknet/issues/179#issuecomment-330047738
But there isn't MobileNet wegihts for classification network, so you will train your own detection network without pre-trained weights: ./darknet detector train obj.data mobile-yolo.cfg