Created by: lqniunjunlper

At first I want to train LR on a large scale dataset using one single machine. In my perspective, paddle only need to maintain parameters and some batch data instead of all training data, so one machine is enough. But i find that Paddle's training process will produce some evaluation results such as cost, auc, etc. and paddle's memory consumption is too large. So i want to know how paddle manage the memory for training and evaluation module during the training process. Besides, how to reduce paddle's memory consumption to minimum? Should i modify the source code to remove some large-memory-consumption module such as global evaluation? Thx~