8943d8bca0
add missing logic to new plugins and accelerator ( #5734 )
...
* add missing logic
* missed imports
* import fixes
* isort
* mv f
* changelog
* format
* move helper to parallel plugin
* d
2021-02-01 13:23:53 -05:00
b3ebc18bcb
Hardware specific parts of Accelerator Refactoring ( #5719 )
...
* add basic accelerator class.
Co-Authored with @awaelchi
* pep8
Co-authored-by: @awaelchi
* add cpu accelerator
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add gpu accelerator
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add tpu accelerator
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add accelerator connector
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add single device training
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add single tpu
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add tpu spawn
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* make on_colab_kaggle utility func
* add basic accelerator class.
Co-Authored with @awaelchi
* pep8
Co-authored-by: @awaelchi
* add cpu accelerator
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add gpu accelerator
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add tpu accelerator
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add accelerator connector
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add single device training
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add single tpu
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add tpu spawn
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* make on_colab_kaggle utility func
* fixes
* move
* yapf
* .
* .
* .
* flake8
* sync accelerator connector changes from dev1.2
* changelog
* fix tpu handling
* tpu
* aval
* yapf
* Update pytorch_lightning/plugins/training_type/tpu_spawn.py
Co-authored-by: chaton <thomas@grid.ai>
* Update pytorch_lightning/accelerators/accelerator_connector.py
Co-authored-by: chaton <thomas@grid.ai>
* Update pytorch_lightning/plugins/training_type/tpu_spawn.py
Co-authored-by: chaton <thomas@grid.ai>
* Update tpu_spawn.py
* Update pytorch_lightning/accelerators/accelerator_connector.py
Co-authored-by: chaton <thomas@grid.ai>
* indentation
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
Co-authored-by: Jirka Borovec <jirka.borovec@seznam.cz>
Co-authored-by: chaton <thomas@grid.ai>
2021-02-01 08:34:59 -05:00
069ae27cef
Accelerator Refactor: Precision Plugins ( #5718 )
...
* add basic accelerator class.
Co-Authored with @awaelchi
* add basic trainign type plugin.
Co-Authored with @awaelchi
* pep8
Co-authored-by: @awaelchi
* update copyright
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add apex_amp
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add mixed base class
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add native amp
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add native amp sharded
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add tpu bfloat
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* add inits
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* Update precision_plugin.py
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
2021-01-31 13:12:02 -05:00
5d239ccd70
Base classes for accelerator refactoring ( #5715 )
...
* add basic accelerator class.
Co-Authored with @awaelchi
* Add base plugin class.
Co-authored with @awaelchi
* add basic trainign type plugin.
Co-Authored with @awaelchi
* add basic precision plugin.
Co-Authored with @awaelchi
* Add missing inits.
Co-authored with @awaelchi
* pep8
Co-authored-by: @awaelchi
* ignore flake8
* coverage omit
* imports in init
* lost
* imports
* flake8
* .
* .
* chlog
* Update pytorch_lightning/plugins/training_type/training_type_plugin.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update pytorch_lightning/plugins/training_type/training_type_plugin.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update pytorch_lightning/plugins/training_type/training_type_plugin.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update pytorch_lightning/plugins/training_type/training_type_plugin.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update pytorch_lightning/plugins/training_type/training_type_plugin.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update pytorch_lightning/plugins/training_type/training_type_plugin.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update pytorch_lightning/plugins/training_type/training_type_plugin.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
Co-authored-by: Jirka Borovec <jirka.borovec@seznam.cz>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
2021-01-30 14:55:28 -05:00
7e2e874d95
Refactor: legacy accelerators and plugins ( #5645 )
...
* tests: legacy
* legacy: accel
* legacy: plug
* fix imports
* mypy
* flake8
2021-01-26 20:04:36 -05:00
2846322f60
fix docs render ( #5610 )
2021-01-25 20:21:00 -05:00
e806bb77fa
Refactor LightningDistributedDataParallel ( #5185 )
...
* add wrapper
* add squeeze
* replace LightningDistributedDP
* update import
* module access
* inputs
* refactor warning
* update
* resolve flake8
* remove old class
* set find unused params to False
* update docstrings
* update docs
* update docs
* add changelog
* deprecation
* rename wrapper -> module
* rename pl_module
* add unit tests
* Revert "add changelog"
This reverts commit 02ec0a6864f4ba2ace3bb6fc6ebc364e1a80ffd7.
* Revert "set find unused params to False"
This reverts commit 8e451515e6ba3227d00f4a5cb63f332cfedb7b30.
Co-authored-by: Ubuntu <thomas@grid.ai>
2021-01-13 14:35:42 -05:00
0f36525e8f
fix/enable - check F401 ( #5201 )
...
* refactor - check F401
* missed
* fix
2020-12-21 10:15:04 +01:00
2c3d43dcb5
Initialize trainer with None in DDPAccelerator ( #4915 )
...
* Initialize trainer with None
* add typing to all accelerators
* resolve imports
* update
* add typing
* removed typo
* update
* Fix formatting and imports in accelerator
Co-authored-by: maxjeblick <maxjeblick@users.noreply.github.com>
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
Co-authored-by: SeanNaren <sean@grid.ai>
Co-authored-by: Roger Shieh <sh.rog@protonmail.ch>
2020-12-10 15:24:44 +01:00
cdbddbe99f
release 1.1.0 ( #5048 )
...
* release 1.1.0
* pep8
2020-12-10 00:52:39 +00:00
ce9179591d
ref: clean config [1/n] add intermediate setters ( #4990 )
...
* add intermediate setters
* show inputs
* fix options
* move
* fix
* less talk
* fix
* talk less
* str
* cases
* rename
Co-authored-by: chaton <thomas@grid.ai>
2020-12-09 14:13:57 -05:00
127454ade2
All gatherwith grads ( #5012 )
...
* all_gather
* ddp
* horovod
* grad tests
* fixed ddp
* ddp fixed, removed tpu, horovod for now
* changelog
* windows fix
* windows fix
* removed batch from ctx
* all_gather
* ddp
* horovod
* grad tests
* fixed ddp
* ddp fixed, removed tpu, horovod for now
* changelog
* windows fix
* windows fix
* removed batch from ctx
* removed code duplication
* merge
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
2020-12-08 23:20:01 +00:00
ee9b3fe574
[feat] pp 1/n ( #5016 )
...
* Added changes for RPC plugin
* Add missing kwargs
* Fix code format
* Loading refactors by introducing is_distributed var, fix optimizer step flow
* Add rpc guard
* Added docstrings and typing
* resolve comments
* Add additional rpc hook, refactor name of exit process hook for clarity
* remove annotation
* Modify behaviour to allow optional return, add test for rpc plugin
* resolve tests
* rename is_ddp_based
* update
* update for windows
* update
* resolve test
* code smell
* Revert back to init_ddp_connection for backwards compat
* Swap to explicit name for property
* Add missing speed parity increase for CI variability, fix call counts for child process
Co-authored-by: tchaton <thomas@grid.ai>
2020-12-08 22:02:10 +00:00
2393474350
[hotfix] ddp + manual_optimisation ( #4976 )
...
* Rely on ddp plugin for blocking sync behaviour, and skip if we're using manual optimization
* debug
* Revert "debug"
This reverts commit ccca6b6b
2020-12-07 19:31:54 +00:00
02152c1729
Simplify optimization Logic ( #4984 )
...
* Rely on ddp plugin for blocking sync behaviour, and skip if we're using manual optimization
* debug
* Revert "debug"
This reverts commit ccca6b6b
2020-12-07 12:55:49 +00:00
12cb9942a1
Tpu save ( #4309 )
...
* convert xla tensor to cpu before save
* move_to_cpu
* updated CHANGELOG.md
* added on_save to accelerators
* if accelerator is not None
* refactors
* change filename to run test
* run test_tpu_backend
* added xla_device_utils to tests
* added xla_device_utils to test
* removed tests
* Revert "added xla_device_utils to test"
This reverts commit 0c9316bb
* fixed pep
* increase timeout and print traceback
* lazy check tpu exists
* increased timeout
removed barrier for tpu during test
reduced epochs
* fixed torch_xla imports
* fix tests
* define xla utils
* fix test
* aval
* chlog
* docs
* aval
* Apply suggestions from code review
Co-authored-by: Jirka Borovec <jirka.borovec@seznam.cz>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: chaton <thomas@grid.ai>
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
2020-12-02 13:05:11 +00:00
404af43cde
5/n: Extract reference model call to plugins/accelerators ( #4773 )
...
* Encapsulate extracting reference model within the plugin to allow custom wrapper logic to live within the plugin/accelerators
* Add missing new lines
* Fix call to accelerator
* Removed double blank
* Use accelerator backend
* Handle case where wrapper has not been initialized within the plugin
* Added basic get model tests, add better typing
* Change model name
* Split GPU/DDP test
* Add stronger typing, skip ddp test on windows
* Fix import
* Fix import in dp
* Fixed PEP8 definition
* Add ddp launcher for ddp testing
* Modify accelerator reference model to property, change name to reflect func
* Revert property as this is incorrect.=
* Revert across accelerators
* Modified name to get_model_from_plugin
* Code review changes, fix issue with dp
* Add verb to function getter
Co-authored-by: chaton <thomas@grid.ai>
2020-11-23 17:21:47 +00:00
45c57600af
Move init_ddp_connection to DDP Plugin ( #4407 )
...
* Move init_ddp_connection to DDP Plugin
* cluster-env
* trainer?
* imports
* Update ddp_plugin.py
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
2020-11-18 15:49:22 -05:00
e7134a9135
Sharded Plugin 2/n: Allow ddp plugin to modify optimizer state saving ( #4675 )
...
* Allow ddp plugin to modify optimizer state saving
* Rely on the accelerator for optimizer states
* Ensure we init the accelerator for the saving function
* Better comment for optim state dump
* Revert "Ensure we init the accelerator for the saving function"
This reverts commit af65effaf9929c0c
2020-11-18 16:38:35 +00:00
4018237c30
[FEAT] Add lambda closure to manual_optimizer_step ( #4618 )
...
* added lambda_closure
* move to types
* add 2 new tests
* make example more complex
* add complex example to doc
* added more tests
* resolve doc
* typo
* update
* update tpu optimizer_step
* Apply suggestions from code review
* Update pytorch_lightning/core/lightning.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* update
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
2020-11-12 19:22:06 +00:00
bacabaebaf
Sharded Accelerator 1/n: Expose clip gradients to plugins via abstract class ( #4639 )
...
* Added abstract precision plugin to expose clip_gradients function, use within accelerator to clip gradients
* Exclude model from override, keep optimizer (needed for sharded clip gradients), add override for O2 support apex
* Fix doc
* Applied codereview changes
* Refactored clip function to encapsulate tpu changes with tpu accelerator. Default to standard clip function for vanilla torch
* Pass correct grad clip val
* Moved var to property
* Apply code review suggestions
2020-11-12 17:18:09 +00:00
7e08b0d710
[bug-fix] DDP and automatic_optimization=False ( #4485 )
...
* resolve bug
* add self._running_manual_optim
* update
* update tests
* update lightning module
* resolve bug
* update tests
* update
* resolve pep8
* update
* replace by `ddp_spawn`
* temporary fix
* update
* update
* move update to training_loop
* make both ddp_spawn
* introduce `manual_optimizer_step`
* update changelog
* added changelog wrong place
* add force_optimizer_step
* update docstring for tests
* update optimizer_step
* update zero_grad
* resolve flake8
* move update into manual_optimizer_step
* add zero_grad
* remove zero_grad tests
* remove manual_backward in AMP, it doesn't help
* update
* loosen tests
* update
* update doc
* add TODO
* Removed unnecessary get model from native amp
* Remove try except with pytest raise
* Add seed, clean up imports, remove try catch to reproduce error
* update code
* update test
* revert back
* formatting
* Update pytorch_lightning/core/lightning.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: SeanNaren <sean@grid.ai>
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
2020-11-10 19:44:51 +00:00
ee35907170
Accelerator docs ( #4583 )
...
* accelerator docs
* accelerator docs
2020-11-08 17:24:41 -05:00
51cc7a89ee
Horovod: fixed early stopping and added metrics aggregation ( #3775 )
...
* Fixed early stopping for Horovod
* Refactored to sync_dist_if_available
* Bump min Horovod version to support hvd.is_initialized
* Changelog
* Added back change for Horovod
* Removed redundant checks for initialization
* Implement metrics gathering for Horovod
* Added test for EvalResult
* Renamed ddp_sync_on_step -> dist_sync_on_step
* Added metric test for Horovod
* Added option pass callable allgather function to metric base class
* Added dist_sync_fn
* Fixed calls to private _sync_dist
* Fixed Horovod test
* Added sync_tensor to the distributed backend
* Skip Windows
* Insert test path
* Removed redundant import
* Updated drone
* Unset HOROVOD_GPU_ALLREDUCE
* Unset
* No cache dir
* No uninstall
* Unset variables
* Uninstall Horovod during initialization
* Replaced more references to ddp_sync_on_step
* Fixed imports
* Fixed attribute
* Added back default
* Lint
* Added back docstring
* Made gather_all_tensors default
* Added whitespace
* Update tests/models/test_horovod.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update pytorch_lightning/metrics/metric.py
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
* Update CHANGELOG.md
Co-authored-by: Teddy Koker <teddy.koker@gmail.com>
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
2020-11-05 12:52:02 -05:00
01ab2a933d
[bug] [docs] Clearer optimizer_step override instructions ( #4455 )
...
* fix
* flags
* remove defaults
2020-11-02 22:13:34 +00:00
102fa9ee7d
[BUGFIX] AMP + Precision unscale grad ( #4441 )
...
* move unscale within Native plugin
* remove gradient tracking from lightning backward
* forgot trainer.fit
* typo
* update
* cleanup
* set to 1.6
* typo
* skip if below 1.6 strict
* update changelog
* remove useless code
* Update tests/plugins/test_amp_plugin.py
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
* Update tests/plugins/test_amp_plugin.py
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
* update changelog
* Update CHANGELOG.md
Co-authored-by: Sean Naren <sean.narenthiran@gmail.com>
Co-authored-by: Jeff Yang <ydcjeff@outlook.com>
2020-11-02 16:36:48 +00:00
5641b266d5
Bug/4319 ddp checkpoint ( #4323 )
...
* Broadcast best model path to ensure we sync with main process + wait for main process to save
* Add barrier call to ensure all processes are in sync
* Added changelog commit
* Move sync of best model path/score to model checkpoint, keep barrier to ensure all processes complete
* Ensure we broadcast as tuple
* Add init check
* Update pytorch_lightning/callbacks/model_checkpoint.py
Co-authored-by: ananthsub <ananth.subramaniam@gmail.com>
* Update pytorch_lightning/callbacks/model_checkpoint.py
Co-authored-by: ananthsub <ananth.subramaniam@gmail.com>
* Removed model checkpoint code, added barrier to trainer to enforce we syncronize and wait for all processes to finish before completing training
* Add barrier within teardown call, removed horovod teardown to inherit from base accelerator
Co-authored-by: ananthsub <ananth.subramaniam@gmail.com>
2020-10-24 16:55:49 -04:00
753362d0a4
enable ddp as a plugin ( #4285 )
...
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
* enable custom ddp plugin
Co-authored-by: chaton <thomas@grid.ai>
2020-10-22 05:15:51 -04:00
0ec4107697
Optimizer closure ( #4190 )
...
* closure for all optimizers
* rename hook and take care of alternating backwards
* add comment
* training_loop_fix
* closure whenever possible
* training_loop
* simple tests that count backward calls
* fix test to work with closure
* remove debugging statement
* better place
* check grads after backward
* start fixing manual optimization
* skip step when result returned by closure was None
* fix gradient clipping test to work with closure
* attribute dict result only for automatic optimization
* adjust backward calls in accelerator
* adjust where to call gradient clipping
* adjust backward calls in tests
* Apply suggestions from code review
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
* pass kwargs to xla optimizer
Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
2020-10-21 19:34:29 +01:00
b45b57cc58
Use `Optional` for arguments set to `None` by default ( #4164 )
...
* Use `Optional` for variables set to `None` by default
* Use `Optional` instead of `Union[None, ...]` for consistency
2020-10-15 23:02:50 +02:00
98eb736496
Added getstate/setstate method for torch.save serialization ( #4127 )
...
* Added getstate/setstate method for torch.save serialization, added additional Optional Typing to results object
* Added tests to ensure torch.save does not fail
* Added flags to ensure compatible ddp cpu environment
* Removed torch version check due to minimum already being 1.3, reduced epochs for speed
* Moved tests to separate file
* Update to accelerator, move to ddp_spawn to prevent hanging ddp
2020-10-13 16:47:23 -04:00
09c2020a93
notices ( #4118 )
2020-10-13 07:18:07 -04:00
b9f2682b7d
clean docs, enable grad clip in manual mode ( #4078 )
...
* docs
* docs
2020-10-11 13:12:35 -04:00
7ffe05a3d1
ref: accelerator names ( #4066 )
...
* ref: accelerator names
* docs
2020-10-11 01:05:14 -04:00
Adrian Wälchli