* Added basic functionality and tests
* Feature parity with old tune search space config
* Convert Optuna search spaces
* Introduced quantized values
* Updated Optuna resolving
* Added HyperOpt search space conversion
* Convert search spaces to AxSearch
* Convert search spaces to BayesOpt
* Added basic functionality and tests
* Feature parity with old tune search space config
* Convert Optuna search spaces
* Introduced quantized values
* Updated Optuna resolving
* Added HyperOpt search space conversion
* Convert search spaces to AxSearch
* Convert search spaces to BayesOpt
* Re-factored samplers into domain classes
* Re-added base classes
* Re-factored into list comprehensions
* Added `from_config` classmethod for config conversion
* Applied suggestions from code review
* Removed truncated normal distribution
* Set search properties in tune.run
* Added test for tune.run search properties
* Move sampler initializers to base classes
* Add tune API sampling test, fixed includes, fixed resampling bug
* Add to API docs
* Fix docs
* Update metric and mode only when set. Set default metric and mode to experiment analysis object.
* Fix experiment analysis tests
* Raise error when delimiter is used in the config keys
* Added randint/qrandint to API docs, added additional check in tune.run
* Fix tests
* Fix linting error
* Applied suggestions from code review. Re-aded tune.function for the time being
* Fix sampling tests
* Fix experiment analysis tests
* Fix tests and linting error
* Removed unnecessary default_config attribute from OptunaSearch
* Revert to set AxSearch default metric
* fix-min-max
* fix
* nits
* Added function check, enhanced loguniform error message
* fix-print
* fix
* fix
* Raise if unresolved values are in config and search space is already set
Co-authored-by: Richard Liaw <rliaw@berkeley.edu>
SAC (both torch and tf versions) are showing issues (crashes) due to numeric instabilities in the SquashedGaussian distribution (sampling + logp after extreme NN outputs).
This PR fixes these. Stable MuJoCo learning (HalfCheetah) has been confirmed on both tf and torch versions. A Distribution stability test (using extreme NN outputs) has been added for SquashedGaussian (can be used for any other type of distribution as well).
* Checkpoint the image-models example
* Update cluster definition
* Fix copyright info
* Use original args
* Checkpoint fixes
* Add README
* Add some missing features
* Format
* Get rid of the unused Namespace class
* Address comments
* Link the imagenet example in docs
* Cleanup
* Fix lint
Why are these changes needed?
Running a worker on head (locally, not as a Ray actor) allows for easier handling of stateful stuff like logging and for easier debugging.
* Update issue templates
* Init fp16
* fp16 and schedulers
* scheduler linking and fp16
* to fp16
* loss scaling and documentation
* more documentation
* add tests, refactor config
* moredocs
* more docs
* fix logo, add test mode, add fp16 flag
* fix tests
* fix scheduler
* fix apex
* improve safety
* fix tests
* fix tests
* remove pin memory default
* rm
* fix
* Update doc/examples/doc_code/raysgd_torch_signatures.py
* fix
* migrate changes from other PR
* ok thanks
* pass
* signatures
* lint'
* Update python/ray/experimental/sgd/pytorch/utils.py
* Apply suggestions from code review
Co-Authored-By: Edward Oakes <ed.nmi.oakes@gmail.com>
* should address most comments
* comments
* fix this ci
* first_pass
* add overrides
* override
* fixing up operators
* format
* sgd
* constants
* rm
* revert
* Checkpoint the basics
* End of day checkpoint
* Checkpoint log-to-head implementation
* Checkpoint
* Add actor-based batch log reporting, currently segfaults
* Work around progress segfault
* Fix some stuff in quicktorch
* Make things more customizable
* Quality of life fixes
* More quality of life
* Move tqdm logic to training_operator
* Update examples
* Fix some minor bugs
* Fix merge
* Fix small things, add pbar to dcgan
* Run format.sh
* Fix missing epoch number for batch pbar
* Address PR comments
* Fix float is not subscriptable
* Add train_loss to pbar by default
* Isolate tqdm code into a handler system
* Format
* Remove the batch_logs_reporter from distributed runner as well
* Check if the train_loss is avaialbale before using it
* Enable tqdm in the dcgan example
* Fix a crash in no-handler trainers
* Fix
* Allow not calling set_reporters for tests
Co-authored-by: Philipp Moritz <pcmoritz@gmail.com>
Co-authored-by: Richard Liaw <rliaw@berkeley.edu>
Co-authored-by: Edward Oakes <ed.nmi.oakes@gmail.com>
Ian Rodney