* 1. Create buddy allocator in each places before NcclBcast the variables
2. Check the memory usage of ALL gpus rather than the first one

* 1. Make NCCLGroupGuard guards only the ncclBcast part, which avoid ncclGroupEnd blocking the exception throwing
2. NOTE the usage of NCCLGroupGuard

* Remove the memory usage check of gpus

* Fix code style

outvar must be create in local scope for prefetch

Remove cuptiFinalize.

* Add mean_iou op.

* Add unitest for mean iou op.

* Add optional collections of confusion matrix and mean_iou.

* Fix cuda kernel.

* Refine code.
1. Merge computing in GPU to two kernel.
2. Use wrong array and correct array instead of confusion matrix.

* Add python api and fix cuda kernel.

* Fix comments.

* Small fix.

* Small fix.

In cupti samples, only cuptiFlush is used.
I can't find any places calling cuptiFinalize and
this API can error out as not_implemented in some
cuda installation.

* fix #11416

* make sgd check tape has been backwarded_

* add error message

* change dynamic to tape

* update readme link

add apis for learning_rate_scheduler

Add merge_ids_op

fix doc generating error

fix errors

MKLDNN layout: Support for convolution operator

Fix nccl dist train bug

Change drop_last defalut value

add url of cuda9.0_cudnn7_avx_mkl

* expose h0 in dynamic_lstm

* update by comment

* update by comment

* h0 to H0

add detection api

fix build on mac

MKLDNN layout: Support for batch norm operator

small fix