Created by: luotao1

根据【Paddle模型性能统计规范】：

1. 计算ips：ips单位在本Repo是images/sec
2. 计算batch_cost：多轮训练1次打印时，batch_cost应求平均值。注意，连续统计，不要遗漏某些轮次。（develop分支已经符合要求）
3. 格式上：batch_cost、reader_cost、ips与模型其它指标共同打印，保留小数点后5位。ips，必须打印！ ips放在整条打印最后

本PR：

1. 增加了ips计算
2. batch_cost从小数点后3位改成5位，同时补充了单位。

• 命令：python tools/train.py -c configs/yolov3_mobilenet_v1_roadsign.yml --eval -o use_gpu=0
• 效果： before: batch_size=8, log_iter=20

2020-09-25 10:47:02,989-INFO: iter: 0, lr: 0.000033, 'loss': '5710.631836', time: 0.000, eta: 0:00:00
2020-09-25 10:58:01,319-INFO: iter: 20, lr: 0.000047, 'loss': '29.493576', time: 32.155, eta: 10:32:22

after: batch_size=8, log_iter=2

2020-09-25 15:10:53,118-INFO: iter: 0, lr: 0.000033, 'loss': '11803.891602', batch_cost: 0.00006 s, eta: 0:00:00, ips: 125203.10448 images/sec
2020-09-25 15:11:54,840-INFO: iter: 2, lr: 0.000035, 'loss': '1264.233521', batch_cost: 20.69376 s, eta: 6:53:11, ips: 0.38659 images/sec
2020-09-25 15:12:57,061-INFO: iter: 4, lr: 0.000036, 'loss': '557.037598', batch_cost: 26.74744 s, eta: 8:53:09, ips: 0.29909 images/sec
2020-09-25 15:13:51,134-INFO: iter: 6, lr: 0.000037, 'loss': '225.754166', batch_cost: 26.11755 s, eta: 8:39:44, ips: 0.30631 images/sec
2020-09-25 15:14:48,563-INFO: iter: 8, lr: 0.000039, 'loss': '79.420753', batch_cost: 26.78939 s, eta: 8:52:12, ips: 0.29863 images/sec
2020-09-25 15:15:44,973-INFO: iter: 10, lr: 0.000040, 'loss': '46.613548', batch_cost: 27.72659 s, eta: 9:09:54, ips: 0.28853 images/sec