Up and Down-Sampling in Training Step

[zhangpengshan]

Some times if not enough or over-enough data with positive tags, up and down sampling are used to improve such situation.

[zhangpengshan]

SampleNegtiveOnly already in norm step;
Add upSampleWeight in train step to up sampling positive records: If set upSampleWeight to 5, then positive will be randomized up to 5 times according to possion distribution.

[zhangpengshan]

By default upSampleWeight is 1.0 and hidden in ModelConfig.json, which means no any up sampling.