dataset sink mode

Merge pull request !8085 from ougongchang/fix_round

Merge pull request !7910 from luopengting/summary

Merge pull request !7534 from ougongchang/feature_mindexplain

The SummaryRecord.add_value() method is extended to record the data of
MindExplain.

Merge pull request !6781 from ZPaC/master-change-prefix-for-server-ckpt

change ImagefolderDV2 name

change ds.transforms.vision to ds.vision

change batch api to match map api more closely

compose op changes

test_pylint

remove compose op from vision, move to transform module, refactor map and batch to use column_order

Merge pull request !5482 from liuyang/md_save_checkpoint

Merge pull request !5389 from ougongchang/fix_summarycollector

- handle collection for multiple trains
- how many tensors to collect when sunk
- change loglevel for get_learning_rate
- update calculation of `max_file_size`
- fix how collect_tensor_freq counting

Merge pull request !3907 from changzherui/mod_ckpt_func_param

- fix writer pool not exit when max_file_size too small
- fix API docs to illustrate `collect_tensor_freq` and `max_file_size`

We collect graph and dataset graph in begin stage before,
If there compile graph fail in GPU, we also collect graph
and dataset graph to summary dir, it will confuse user.

So we collect graph and dataset graph in step end stage now,
If there compile graph fail, we will not collect graph and dataset
graph.

When `collect_tensor_freq` is specified as `None`,
the `collect_tensor_freq` would be auto calculated.

The previous behavior is to collect at most 50 steps,
now changing to 20

1. collect the origin network in model, and set it to cb_params
2. collect the origin network name in SummaryCollector
3. Update the SummaryCollector API Doc

The useage is error when collect custom lineage data.

Add more detail note for SummaryCollector and SummaryRecord,
else if it is used not right, some proplem will be caused.

Merge pull request !2718 from chenzhongming/master