search

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).
MIT License
1.73k stars 784 forks source link

Calculate mAP/PR curves and view both losses training/validation at same time #173

Open PythonImageDeveloper opened 6 years ago

PythonImageDeveloper commented 6 years ago

Hi, I want to Calculate mAP/PR curves and view both losses training/validation at same time , what do i do ? your work is excellent but why didn't calculate those metrics ? in my opinion ,if we can to plot both losses training/validation at the same time of training duration , we can decide to when to stop the training procedure , and we can to know when the Early stopping occurred. I've always been questioned why in the deep frameworks didn't use any scripts about view of both losses. in my opinion , if didn't have such plots ,Where to know When should stop the training procedure?

experiencor commented 6 years ago

@zeynali You may refer to https://github.com/experiencor/basic-yolo-keras/issues/27 for the code to compute mAP. It's lots of code to write. If I remember it correctly, https://github.com/tensorflow/models/tree/master/research/object_detection allows you to visualize mAP progress in TensorBoard.

RichardSieg commented 6 years ago

Hey. You should consider using https://pypi.python.org/pypi/plotille for plotting the curve. Internally, we used the algorithm proposed in the very well written code for RetinaNet: https://github.com/fizyr/keras-retinanet/blob/master/keras_retinanet/utils/eval.py