1. Introduction In this post, I would like to compare the loss functions used in different one-shot object detection methods, YOLO, SSD, and RetinaNet. One-shot object detection methods train the model on more than thousands grids with different scale, but the number of objects in one image is much... [Read More]

In this post, I would like to review the development of optimization methods for deep neural network (DNN) and share suggestions to use optimizers. What you can find: A brief review of the popular optimizers from the an intuitive perspective. The disadavantage of the popular adaptive optimizer, Adam.... [Read More]

Tuning the hyper-parameters of a deep learning (DL) model by grid search or random search is computationally expensive and time consuming. This technical report gives several practical suggestions and steps to choose the optimal hyper-parameters. Some prior knowledge to fully understand this technical report: Overfitting/Underfitting Learning Rate (LR) Batch... [Read More]