Skip to content

Add broadcast operators#17503

Merged
chengduoZH merged 7 commits intoPaddlePaddle:developfrom
chengduoZH:feature/dygraph_param_bcast
May 24, 2019
Merged

Add broadcast operators#17503
chengduoZH merged 7 commits intoPaddlePaddle:developfrom
chengduoZH:feature/dygraph_param_bcast

Conversation

@chengduoZH
Copy link
Contributor

@chengduoZH chengduoZH commented May 20, 2019

Demo: https://gist.github.com/chengduoZH/1a34d84c0672a46984bd17c16b6d9dd7
This PR adds broadcast for multi-process. And it could be used in dynamic graph to broadcast parameters.

test=develop
@chengduoZH chengduoZH force-pushed the feature/dygraph_param_bcast branch from fd952d0 to 3f7c77d Compare May 20, 2019 15:03
test=develop
@chengduoZH chengduoZH force-pushed the feature/dygraph_param_bcast branch from 3f7c77d to e71d181 Compare May 20, 2019 23:24
test=develop
@chengduoZH chengduoZH force-pushed the feature/dygraph_param_bcast branch 2 times, most recently from 0ddf31a to ce53b58 Compare May 21, 2019 15:49
test=develop
@chengduoZH chengduoZH force-pushed the feature/dygraph_param_bcast branch from ce53b58 to 2b3b07c Compare May 21, 2019 23:51
test=develop
@chengduoZH chengduoZH changed the title [WIP] Add broadcast operators Add broadcast operators May 23, 2019
@chengduoZH chengduoZH requested a review from JiabinYang May 23, 2019 04:06
@chengduoZH chengduoZH force-pushed the feature/dygraph_param_bcast branch from 626a5ed to 802bbf1 Compare May 23, 2019 04:12
test=develop
@chengduoZH chengduoZH force-pushed the feature/dygraph_param_bcast branch from 802bbf1 to 5d641ba Compare May 23, 2019 04:18
Copy link
Contributor

@JiabinYang JiabinYang left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we have an unit test here?

@chengduoZH
Copy link
Contributor Author

can we have an unit test here?

There have two unit tests for dynamic graph training by multi-process, i.e. test_parallel_dygraph_mnist.py, test_parallel_dygraph_se_resnext.py

@luotao1
Copy link
Contributor

luotao1 commented May 23, 2019

test_parallel_dygraph_mnist.py, test_parallel_dygraph_se_resnext.py

These tests are model-level, should we add op-level tests?

test=develop
@chengduoZH chengduoZH force-pushed the feature/dygraph_param_bcast branch from f7825d0 to 4b43823 Compare May 23, 2019 08:57
@chengduoZH
Copy link
Contributor Author

@luotao1 We found that the unit test of ncclBroadcast is random failed, and this is the issue #9540, I will analysis the reason and add a unit test in next PR.

Copy link
Contributor

@JiabinYang JiabinYang left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@chengduoZH chengduoZH merged commit b5f4d5e into PaddlePaddle:develop May 24, 2019
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants