Pytorch-yolov3: yolov3.weight and darknet53.conv.74
what's the difference? could anyone tell me, I'm a little confuse
qilong-zhang
All 5 comments
Other than this confusion over the difference, have you been able to run the test.py without any issues?
KtK99
on 11 Jun 2019
AFAIK darknet53.conv.74.weights contains the weights for the DarkNet network* originally trained for classification on the ImageNet dataset, which is used as the pre-trained feature extractor (backbone) for YOLOv3. To use this for detection the additional weights which are only present in the YOLOv3 network are randomly initialized prior to training.
yolov3.weights contains the weights for the full YOLOv3 network after being trained for detection on the COCO dataset (after being initialised with the darknet53.conv.74 weights).
*darknet53.conv.74.weights refers to the fact that only the weights for the convolutional layers are included, excluding the weights for the final fully connected layer which outputs the class probabilities for classification. There is a separate file called darknet53.weights which includes _all_ the weights.
david-macleod
on 18 Jun 2019
how does it affect the training?
shouryasimha
on 31 Mar 2020
@david-macleod I'm having issues using yolov3.weights that have been pretrained on the COCO dataset. I don't want to use darknet53.conv.74.weights since they were pretrained on imagenet's dataset. Using windows cmd and the correct paths in the below command line, the exe does not start training at all: darknet.exe detector train x64/data/obj.data x64/yolov3-openimages.cfg x64/yolov3-openimages.weights
USNA2014
on 6 Apr 2020
If you used yolov3.weights, the model would not train. I have solved the reason why. Please check out the link: https://github.com/AlexeyAB/darknet/issues/5012#issuecomment-609596129
USNA2014
on 6 Apr 2020
Related issues
codeyogi911
路
4Comments
mwharton3
路
3Comments
alexandru-dinu
路
3Comments
ghost
路
3Comments
zhaowt61
路
4Comments
Most helpful comment
AFAIK darknet53.conv.74.weights contains the weights for the DarkNet network* originally trained for classification on the ImageNet dataset, which is used as the pre-trained feature extractor (backbone) for YOLOv3. To use this for detection the additional weights which are only present in the YOLOv3 network are randomly initialized prior to training.
yolov3.weights contains the weights for the full YOLOv3 network after being trained for detection on the COCO dataset (after being initialised with the darknet53.conv.74 weights).
*darknet53.conv.74.weights refers to the fact that only the weights for the convolutional layers are included, excluding the weights for the final fully connected layer which outputs the class probabilities for classification. There is a separate file called darknet53.weights which includes _all_ the weights.