Some partial supports for READ-only inplaced operations (Re-based) #1103
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sharing memories for Reshape(f/b), NoBackprop(f), FlipGradient(f).
* Fixed pernicious bug in autobatching * Reverted to async and added stream
* Add NoneType assert for list arguments in _dynet.pyx. * remove additional assert
* fix same_dims check for cmult fwd * improved same_dims checks * Removed error in affine transform arg check * Some refactoring of broadcasting * Better profiling * Updates for cwise * Updated broadcasts * Fixed autobatch profile for csum
* Added more information in case of memory allocation error * Added two more lines for clarification * Added deviceID in cudaErrorMemoryAllocation error message
Trivial fix to make doxygen url work
some sbt versions ignored the options if they were in the wrong place
* Add pool memory info for profiling. * update * Add profile info when out of memory. * move show_pool_mem_info out of CG class and add it into low-level control code.
…only, and single-batch)
* make {Nary, Unidirectional}TreeLSTMBuilder non-abstract again
Commit bf29c18 added set_h_impl to
RNNBuilder as an abstract method and the TreeLStmBuilders became
abstract. This was fixed for the BidirectionalTreeLSTMBuilder but not
the other two. This commit moves the the stub throwing a runtime error
from BidirectionalTreeLSTMBuilder to the TreeLSTMBuilder super class.
* TreeLSTMs: add nodes in arbitrary order; swig bindings
Main feature:
TreeLSTMs had to be constructed with the nodes in the tree added by
their index. This is not practical if the nodes are already
sequentially ordered with the child nodes not being first, e.g. when
working with dependency trees. If set_num_elements is called, a fixed
set of node space is reserved and nodes can be added in any order as
long as children are added first.
The old behavior still works as usual.
Also:
- Add documentation to treelstm C++ code
- add swig bindings
- add scala wrapper for uni- and bidirectional tree lstm
* scala uni- and bidirectional tree lstm wrapper
* added complex structures page to docs * renamed complex structures page
* Removed devices.h from .h files * Made C++ compile * Fix python * Fixeed examples and tests * Fixed compile on Linux and CUDA * Fixed bug on Windows? * Cleaned headers of examples
…tricter grad check
* Run Travis CI on release tags * Create sdist before build to avoid unnecessary files * Fix replacement in .appveyor.yml
* Enhanced implmentation of --dynet-gpu option in python end * fix copy list bug
…to zzsfornlp-read_inplace0
|
OK, I finally got around to taking a look. First, thanks again for contributing this! I did a few changes to ensure that things don't break with autobatching. The biggest one is that the behavior of the nodes themselves are not modified, but forward or backward is just skipped in the executor. This is useful because now the executor is solely in charge of handling inplacing properly, and if inplacing is not supported things will work as they did before. I'm going to do a few more tests and also make sure that the memory profiling code is working properly, then probably merge this. |
|
Thanks a lot for your help! |
|
#1156 was merged! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a re-based PR of #929 basing on the current master branching, sorry that I have not been following the new commits and changes between that PR and the current point. Thus about this I might need some help for checking if there will be any logical conflicts. Thanks!
Here is some notes about this PR: