* Remove race between local scheduler disconnecting and global scheduler
assigning a task
* Fix number of workers started in component failures test
* Fix race between global scheduler retrying a task assignment and monitor
cleaning up task table. The global scheduler should only retry the task
assignment if the local scheduler is still alive.
* Clean up task_table_update callback if failure
* Look up current local scheduler mapping when retrying actor task submission
* Log warning if no subscribers received a task table update
* Clean up database handle memory in local scheduler
* In setup.py, move cython to setup_requires and move setuptools_scm to setup_requires.
* Add back pip install of cython when building mac wheels.
* Revert changes to setuptools_scm.
* Check that the correct number of Linux wheels are produced.
* Add back pip install cython when building linux wheels.
* Code for Supporting Shared Models
Running (with vnet modification) - needs to be tested for performance
Summaries
Small refactoring + generalized to more domains
Small fix for jenkins
Linting
linting
Addressing changes
Addressing changes
Update envs.py
Addressing changes
convnet
Merge - new model
final touches
final linting
Changing iterations back
removed extra change
changes for fast experimentation
changes to enable a2c
TEMP FOR DEBUGGING
ContinuousActions - Still doesn't work
InvertedPendulum trains with 8 workers - k=200
huber loss
Maxes for InvertedPendulum-v1 - 16w,200steps
temp: working with a2c
Back to shared model
more fixes
small
nit
LSTM to shared models
need to fix last_features
tuning pong
Best record for hitting 0 - with k=16,n=20
nit
a2cremoval
remove A2c reference and nits
nit
removed a2c vestiges
removing a2c
removing example.py
Linting
nit
* Linting + Removing vestigal code
* Final Touches
* nits
* rerun travis
* make information available for GAE
* buggy version of GAE estimator
* fix
* add more logging and reweight losses
* fix logging
* fix loss
* adapt advantage calculation
* update gae
* standardize returns
* don't normalize td lambda ret
* fix
* don't standardize advantages
* do standardization earlier
* different standardization
* initializer
* drop into the debugger
* fix tensorflow broadcasting bug
* vf clipping
* don't standardize tdlambdaret
* different standardization
* use huber loss for value function
* refactor -- first half
* it runs
* fix
* update
* documentation
* linting and tests
* fix linting
* naming
* fix
* linting
* fix
* remove prefix madness
* fixes
* fix
* add value function example
* fix linting
* remove newline
* Pass DPYTHON_EXECUTABLE into cmake for arrow and for ray.
* Add cython to setup.py install_requires.
* Revert custom code for finding python in cmake.
* Correctly find arrow on CentOS.
* In cmake, don't find PythonLibs, just find PYTHON_INCLUDE_DIRS.
* Fix typo.
* Do not use boost shared libraries when building arrow.
* Add six to the setup.py install_requires because it is needed by pyarrow.
* Don't link numbuf against boost_system and boost_filesystem.
* Compile boost when we are on Linux.
* Make numbuf find the correct boost libraries.
* Only use find_package Boost on Linux, suppress output when building boost.
* Changes to wheel building scripts, install cython in mac script.
* Compile flatbuffers ourselves on Linux and pass it in when compiling Arrow.
* Clean up build_flatbuffers.sh and build_boost.sh scripts a little.
* Install cython when building linux wheel.
* Replace pending_tasks utarray with vector.
* Replace local_schedulers vector with unordered_map.
* Replace object info table with unordered_map.
* Replace local_scheduler_plasma_map and plasma_local_scheduler_map with unordered maps.
* Remove unnecessary includes.
* Fix linting.
* Bug fixes.
* Add function for computing the amount of data for a task that wouldn't have to be shipped because it is already accessible to a local scheduler.
* Small cleanups.
* Embedded timeline
* Yeah
* Fixed arrows not showing up.
* Fixed arrows not showing up, and added check boxes for the kinds of dependencies that should be included in the trace.
* first
* Fixes
* Fixed typo in comments, added more comments. fixed linting.
* Added more comments.
* Formatting.
* fixes
* Fixed state.py linting.
* Fixed ui.py linting errors.
* Fixed linting errors.
* Renamed task dependencies and included instructions for viewing arrows.
* Fixed according to PR comments.
* Fixed bug.
* Undid changes to metadata blocks.
* Fixes according to comments.
* Fixed linting.
* Fixed linting.
* NOQA keyword added to link line.
* adding support for the user-interpretable label(UIR)
* more plumbing for num_uirs further upstream; set to infty when specified on cmd line
* pass default num_uirs for actors; update GlobalStateAPI
* support num_uirs in ray.init()
* local scheduler resource accounting: support num_uirs; prep for vectorized resource accounting
* global scheduler test updated
* Fix bug introduced by rebase.
* Rename UIR -> CustomResource and add test.
* Small changes and use constexpr instead of macros.
* Linting and some renaming.
* Reorder some code.
* Remove cpus_in_use and fix bug.
* Add another test and make a small change.
* Rephrase documentation about feature stability.
* Rough pass at installing pyarrow along with Ray.
* Remove hardcoded path and try to find correct path automatically.
* Add print.
* Fix linting.
* Copy pyarrow files to a location that we manually add to python path in order to avoid interfering with pre-existing pyarrow installations.
* Move call to build.sh back into build_ext in setup.py.
* Ignore some linting errors.
* Fix problem in which pyarrow files to copy were listed before they were built.
* Fix tests by importing ray before pyarrow.
* Initial testing of checkpointing functions.
* Save checkpoints in Redis.
* Pipe checkpoint_interval through remote decorator.
* Add a test.
* Small cleanups.
* Submit dummy tasks when reconstructing tasks before the most recent tasks so that we don't end up reconstructing the arguments for those tasks.
* Remove old checkpoints to save space.
* Fix linting.
* Tell cmake which python to use when building arrow.
* Pass different path into cmake when building arrow so that cmake finds the right python.
* Add correct python executable to PATH when running cmake for ray.
* batch small rollouts together
* implement minimum number of samples for each task
* add total time
* fix linting
* style
* fix
* factor out parameters and document stuff
* add rollout batchsize
* address comments
* linting
* small fix
