Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

[Sparse-Gluon] embedding with sparse grad #10924

Merged
merged 8 commits into from
May 16, 2018

Conversation

eric-haibin-lin
Copy link
Member

Description

Add sparse_grad option to hybrid embedding block. This block doesn't have optimizations for distributed training.

Manually checked word_language_model example with sparse_grad=True (without grad clipping). The result is exactly the same as the original implementation.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

  • The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
  • Changes are complete (i.e. I finished coding on this PR)
  • All changes have test coverage:
  • Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
  • Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
  • Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
  • Code is well-documented:
  • For user-facing API changes, API doc string has been updated.
  • For new C++ functions in header files, their functionalities and arguments are documented.
  • For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
  • Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
  • To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

  • Feature1, tests, (and when applicable, API doc)
  • Feature2, tests, (and when applicable, API doc)

Comments

  • If this change is a backward incompatible change, why must this change be made.
  • Interesting edge cases to note here

arg_arrays = {}
contains_sparse = False
for param in self._params:
arg_arrays[param.name] = param.data(self._contexts[0])
Copy link
Member

Choose a reason for hiding this comment

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

did you try to do step on parameter with deferred initialization? what message did you get?

Copy link
Member Author

Choose a reason for hiding this comment

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

Yes I tried the language model example. param.data() was called in the previous implementation (line 113)

if param._grad_stype != 'default':
contains_sparse = True
# update_on_kvstore is set to False by the user
if self._update_on_kvstore is False:
Copy link
Member

Choose a reason for hiding this comment

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

if not self._update_on_kvstore

Copy link
Member Author

Choose a reason for hiding this comment

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

if self._update_on_kvstore is None, I don't need to throw the err

Copy link
Member

Choose a reason for hiding this comment

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

OK

@eric-haibin-lin eric-haibin-lin changed the title [Sparse-Gluon] [WIP] embedding with sparse grad [Sparse-Gluon] embedding with sparse grad May 15, 2018
@@ -114,6 +116,11 @@ def __init__(self, name, grad_req='write', shape=None, dtype=mx_real_t,
self.wd_mult = wd_mult
self.grad_req = grad_req
self.init = init
grad_stype = 'default' if grad_stype is None else grad_stype
Copy link
Contributor

Choose a reason for hiding this comment

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

why not use 'default' as argument default value

contains_sparse = True
# update_on_kvstore is set to False by the user
if self._update_on_kvstore is False:
raise RuntimeError("Cannot set update_on_kvstore when sparse gradients "
Copy link
Contributor

Choose a reason for hiding this comment

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

include parameter name in message

@@ -390,13 +390,14 @@ class Embedding(HybridBlock):
- **out**: N-D tensor with shape: `(x1, x2, ..., xN-1, output_dim)`.
"""
def __init__(self, input_dim, output_dim, dtype='float32',
weight_initializer=None, **kwargs):
weight_initializer=None, sparse_grad=False, **kwargs):
Copy link
Contributor

Choose a reason for hiding this comment

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

documentation for argument

Copy link
Member Author

Choose a reason for hiding this comment

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

Added

@piiswrong piiswrong merged commit b2ccd34 into apache:master May 16, 2018
jinhuang415 pushed a commit to jinhuang415/incubator-mxnet that referenced this pull request May 29, 2018
* draft

* updat test

*  fix kvstore

* fix lint

* fix test

* add proper error msg

* CR comment
rahul003 pushed a commit to rahul003/mxnet that referenced this pull request Jun 4, 2018
* draft

* updat test

*  fix kvstore

* fix lint

* fix test

* add proper error msg

* CR comment
zheng-da pushed a commit to zheng-da/incubator-mxnet that referenced this pull request Jun 28, 2018
* draft

* updat test

*  fix kvstore

* fix lint

* fix test

* add proper error msg

* CR comment
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants