WongKinYiu / PyTorch_YOLOv4

for custom dataset, usually i will split the dataset into development set (training set + validation set) and testing set.
training set is used to training and validation set is used to choose hyperparameters (including hyp, epoch, ...).
and finally use development set with chosen hyper-parameters for training final model.

Originally posted by @WongKinYiu in #278 (comment)

I am curious that, then in the final phase (when you use both train and val), how do you know you model performance (how to decide the "best" time to save a model)?

Plus, train.py needs validation to save model.

Could it be setting "--notest"?

PyTorch_YOLOv4/train.py

Line 469 in eb5f166

    
           parser.add_argument('--notest', action='store_true', help='only test final epoch')

How to use development set using "train.py"