This PR introduces single-node fault tolerance for Tune.
## Previous behavior:
- Actors will be restarted without checking if resources are available. This can lead to problems if we lose resources.
## New behavior:
- RUNNING trials will be resumed on another node on a best effort basis (meaning they will run if resources available).
- If the cluster is saturated, RUNNING trials on that failed node will become PENDING and queued.
- During recovery, TrialSchedulers and SearchAlgorithms should receive notification of this (via `trial_runner.stop_trial`) so that they don’t wait/block for a trial that isn’t running.
Remaining questions:
- Should `last_result` be consistent during restore?
Yes; but not for earlier trials (trials that are yet to be checkpointed).
- Waiting for some PRs to merge first (#3239)
Closes#2851.
## What do these changes do?
Fix the misleading comments in code for:
- `EPISODES_THIS_ITER`
- `EPISODES_TOTAL`
Had noted it before and planned to fix it along with some other changes but seemed very relevant to stay next to #3058 so sending this now.
A fix to an example for tune (`python/ray/tune/examples/pbt_tune_cifar10_with_keras.py`) where the hyperparameters for the optimizer, learning rate and decay, were not being passed into the optimizer.
This means that the current optimizer uses default values for the hyperparameters no matter the config.
Add new search algorithm (genetic) along with the base framework of the searcher (which performs some basic jobs such as logging, recording and organizing in our project).
Note that this is the initial commit. In the following days, we will add example, UT, and other refinements.
Adds the ability for trainables to reset their configurations during experiments. These changes in particular add the base functions to the trial_executor and trainable interfaces as well as giving the basic implementation on the PopulationBasedTraining scheduler.
Related issue number: #2741
It's possible to configure PPO in a way that ends up discarding most of the samples (they are treated as "stragglers"). Add a warning when this happens, and raise an exception if the waste is particularly egregious.
* Added checkpoint_at_end option. To fix#2740
* Added ability to checkpoint at the end of trials if the option is set to True
* checkpoint_at_end option added; Consistent with Experience and Trial runner
* checkpoint_at_end option mentioned in the tune usage guide
* Moved the redundant checkpoint criteria check out of the if-elif
* Added note that checkpoint_at_end is enabled only when checkpoint_freq is not 0
* Added test case for checkpoint_at_end
* Made checkpoint_at_end have an effect regardless of checkpoint_freq
* Removed comment from the test case
* Fixed the indentation
* Fixed pep8 E231
* Handled cases when trainable does not have _save implemented
* Constrained test case to a particular exp using the MockAgent
* Revert "Constrained test case to a particular exp using the MockAgent"
This reverts commit e965a9358ec7859b99a3aabb681286d6ba3c3906.
* Revert "Handled cases when trainable does not have _save implemented"
This reverts commit 0f5382f996ff0cbf3d054742db866c33494d173a.
* Simpler test case for checkpoint_at_end
* Preserved bools from loosing their actual value
* Revert "Moved the redundant checkpoint criteria check out of the if-elif"
This reverts commit 783005122902240b0ee177e9e206e397356af9c5.
* Fix linting error.
* + Compatibility fix under py2 on ray.tune
* + Revert changes on master branch
* + Use default JsonEncoder in ray.tune.logger
* + Add UT for infinity support
A bunch of minor rllib fixes:
pull in latest baselines atari wrapper changes (and use deepmind wrapper by default)
move reward clipping to policy evaluator
add a2c variant of a3c
reduce vision network fc layer size to 256 units
switch to 84x84 images
doc tweaks
print timesteps in tune status
This PR introduces the following changes:
* Ray Tune -> Tune
* [breaking] Creation of `schedulers/`, moving PBT, HyperBand into a submodule
* [breaking] Search Algorithms now must take in experiment configurations via `add_configurations` rather through initialization
* Support `"run": (function | class | str)` with automatic registering of trainable
* Documentation Changes
Richard Liaw