search

qqwweee / keras-yolo3

A Keras implementation of YOLOv3 (Tensorflow backend)
MIT License
7.14k stars 3.45k forks source link

Can I use this repo for models Tiny models created in Darknet? #407

Open JakupGuven opened 5 years ago

JakupGuven commented 5 years ago

I'm training custom models using the Darknet framework using the YOLOv2 and YOLOv3 tiny architectures. When I convert these models to Keras format with your repo @qqwweee and subsequently use coremltools (made by Apple to convert keras and other formats to Apple formats) it doesn't work. I've managed to convert YOLO keras models to CoreML before using YAD2K.

TeMob commented 5 years ago

I'd like to know the same. I've used this repo to convert from tiny-yolov3 to keras to coreML in the past, which worked well. But with my newest project I can't get it to work either. The model in Darknet detects everything, but after the conversion, it isn't able to detect a single object.

JakupGuven commented 5 years ago

@TeMob are you using Darknet by AlexeyAB? I found that they had switched from saving weights as 32-bit unsigned integers to 64 bit ones which messed up the CoreML model, by changing it back I solved it.

TeMob commented 5 years ago

@JakupGuven Yes I am. That's interesting! Where would I go change this? Thanks

JakupGuven commented 5 years ago

@TeMob You'll find everything here: https://github.com/AlexeyAB/darknet/issues/2922#issuecomment-482811870