Created by: lcy-seso

Below is the training log generated by paddle v2 API:

[INFO 2017-07-20 08:13:54,714 train.py:70] Pass 133, Batch 650, Cost 35.525531, {'classification_error_evaluator': 0.7351384162902832}
[INFO 2017-07-20 08:13:59,905 train.py:70] Pass 133, Batch 655, Cost 35.517285, {'classification_error_evaluator': 0.7425796985626221}
[INFO 2017-07-20 08:14:05,056 train.py:70] Pass 133, Batch 660, Cost 35.666832, {'classification_error_evaluator': 0.7405670285224915}

• The train interface only prints cost and evaluating result on current training batch, but no information is provided on average costs and evaluations up to now.
• This is convenient for plotting the learning curve but is not convenient for model selection.
• Previously PaddlePaddle does provide both average costs and costs over current training batch.
• Though users can calculate the average training cost themselves, this is a very basic metric used in machine learning task. I think PaddlePaddle should implement it instead of leaving it to the users.