Created by: pkuyym

测试采用实际音频数据，共run了10个batch，batch size=32，去掉异步和buffer，结论如下：

1. 最主要热点是trans add delta，这个占用了超过80%的时间
2. feeding操作耗时很少，可以暂不考虑优化

主函数耗时分析：

1. 热点是data processing
2. feeding操作耗时很少，可以排除掉这个风险

Total time: 109.838 s
File: profile_sync_reader.py
Function: profile_reader at line 33

Line #      Hits         Time  Per Hit   % Time  Line Contents
==============================================================
    33                                           @profile
    34                                           def profile_reader(epoch_num, batch_size):
    35         2         10.0      5.0      0.0      for epoch_id in xrange(epoch_num):
    36         1          4.0      4.0      0.0          for batch_id, one_batch in enumerate(
    37        11   95392189.0 8672017.2     86.8                  data_reader.batch_iterator(batch_size, batch_size)):
    38        11      46440.0   4221.8      0.0              (bat_feature, bat_label, lod) = one_batch
    39        11      83802.0   7618.4      0.1              res_feature.set(bat_feature, place)
    40        11        452.0     41.1      0.0              res_feature.set_lod([lod])
    41        11       1048.0     95.3      0.0              res_label.set(bat_label, place)
    42        11         90.0      8.2      0.0              res_label.set_lod([lod])
    43        11   14309543.0 1300867.5     13.0              time.sleep(1.3)
    44        11         61.0      5.5      0.0              if batch_id > 9:
    45         1       4296.0   4296.0      0.0                  break