search

Adamdad / keras-YOLOv3-mobilenet

I transfer the backend of yolov3 into Mobilenetv1，VGG16，ResNet101 and ResNeXt101
MIT License
563 stars 166 forks source link

How you used MobileNet model? #40

Open meet-minimalist opened 5 years ago

meet-minimalist commented 5 years ago

Have you used the pretrained MobileNet model and construct yolo layers on top? And used pretrained MobileNet weights for Mobilenet initialization?

Or you just took the architecture of MobileNet and add yolo layers and started training from scratch using voc or coco?

ZH-Lee commented 5 years ago

I used MobileNet weights for MobileNet initialization, and feed VOC2007 datasets into network. But, i got train loss non-convergence（train loss : 30, val loss:26），whaterver i decrease learning rate, i can't got train loss convergence.

meet-minimalist commented 5 years ago

@ZH-Lee so what you did to make it converge?

ZH-Lee commented 5 years ago

@ZH-Lee so what you did to make it converge?

i didn't make it converge. maybe the datasets is small??? I haven't solve the problem.

meet-minimalist commented 5 years ago

@ZH-Lee ohk. 👍

ruoruo6 commented 5 years ago

I used MobileNet weights for MobileNet initialization, and feed VOC2007 datasets into network. But, i got train loss non-convergence（train loss : 30, val loss:26），whaterver i decrease learning rate, i can't got train loss convergence.

How do you train with VOC dataset? just transfer voc format to coco in this code? i attemp to do it but the map is much low than yolov2 on voc. Why ?