model_body.load_weights(weights_path, by_name=True) is not taking properly yolov3.h5 weights.

[chulminkw]

Hello,

It seems like yolov3.h5 weights is not loaded into the model_body if model_body.load_weights() doesn't have skip_mismatch=True parameter value.

yolov3.h5 was made by python tools/model_converter/convert.py cfg/yolov3.cfg weights/yolov3.weights weights/yolov3.h5

In line 239~240 from yolo3.model.py if weights_path: model_body.load_weights(weights_path, by_name=True) #, skip_mismatch=True)

after I changed into skip_mismatch=True it was successfully loaded.

[vse-motec]

Hello chulminkw, are you using the setup with 80 classes? The yolov3.h5 is created from yolov3.weights (as you did) which is trained on COCO with 80 classes as output. If you change the number of classes in your model then you need to use skip_mismatch=True. Otherwise (i.e. with 80 classes) the model should be loaded correctly. Please check this in your setup and report your findings.

[chulminkw]

Hello vse-motec Thanks for the info. I thought skip_masmatch should be True to work with yolo models which don't have 80 classes. That way, layers other than prediction layers would have weights with yolov3.h5. Currently, it gives an error message without loading weights if the model doesn't have 80 classes.

[vse-motec]

I thought skip_masmatch should be True to work with yolo models which don't have 80 classes. That way, layers other than prediction layers would have weights with yolov3.h5.

This is correct. Maybe my answer was not that clear.

If you have a model with 80 classes your model should load correctly even without skip_mismatch=True. If you change the number of classes e.g. for transfer learning, then use skip_mismatch=True. In this case the prediction layers will not be loaded, which is not bad, since you will transfer learn them anyway.

[vse-motec]

@chulminkw please close this issue if you feel that your issues have been resolved

[chulminkw]

Thanks vse-motec