execution & evaluation

execution & evaluation

The related content of execution and evaluation is limited and easy to understand, so we just make a brief introduction here. For more details, you can read the code.

execution

The main function of executor is to organize the whole training process: train&val, weights save&load, log print&save and more. Be aware that, we adopt DataParallel for multi-GPU training.

evaluation

We provider ways of online evaluating models via subclassing base evaluator. Here, a COCO evaluator is implemented for reference. For WIDERFACE and TT100K, we evaluate model offline by using the tools provided by authors.