I am Feiyue Xie, a student studying in University of Science and Technology of China, who uses github to practice programming techniques.
Since I found some issues in keras, mxnet, gluon-cv, jittor, and rust, I learned to write PRs and become their contributer. Although most of the PRs are not huge change and most of them are easy-to-fix performance problem, there is something important change which is worth mention.
In old versions of MXNet (<=1.8.0, don't know whether it affects pytorch or tensorflow), execute data_iter_prefetch=iter(data_iter) after the data_iter is defined and data_iter_prefetch is consumed may speed up the calculation. without such execution may slow down the training process especially you're using some heavy image ppre-processing techniques like autoaugment in a small dataset (e.g., CIFAR-100).
Older version of MXNet prefetch data only after the __iter__() method is called.
In MXNet 1.x.0 (x>=9) with auto_reload=True(not the default setting) or MXNet 2.x with nopython mode, the prefetch is done as expected.