GitLab

Reviewed By: zhanghang1989

Differential Revision: D27783989

fbshipit-source-id: f05c11e396a2f62366721b365929b29f05d5bc02

Synchronize PyTorchLightning/pytorch-lightning (revision 0b843848@master) to github/third-party/PyTorchLightning/pytorch-lightning

Summary:
### New commit log messages
## [UnReleased] - 2021-MM-DD

### Added

- Added more explicit exception message when trying to execute `trainer.test()` or `trainer.validate()` with `fast_dev_run=True` ([#6667](https://github.com/PyTorchLightning/pytorch-lightning/pull/6667))

- Added `LightningCLI` class to provide simple reproducibility with minimum boilerplate training cli. ([#4492](https://github.com/PyTorchLightning/pytorch-lightning/pull/4492))

- Trigger warning when non-metric logged value with multi processes hasn't been reduced ([#6417](https://github.com/PyTorchLightning/pytorch-lightning/pull/6417))

- Added `gradient_clip_algorithm` argument to Trainer for gradient clipping by value ([#6123](https://github.com/PyTorchLightning/pytorch-lightning/pull/6123)).

- Added a way to print to terminal without breaking up the progress bar ([#5470](https://github.com/PyTorchLightning/pytorch-lightning/pull/5470))

- Added support to checkpoint after training steps in `ModelCheckpoint` callback ([#6146](https://github.com/PyTorchLightning/pytorch-lightning/pull/6146))

- Added `checkpoint` parameter to callback's `on_save_checkpoint` hook ([#6072](https://github.com/PyTorchLightning/pytorch-lightning/pull/6072))

- Added `RunningStage.SANITY_CHECKING` ([#4945](https://github.com/PyTorchLightning/pytorch-lightning/pull/4945))

- Added `TrainerState.{FITTING,VALIDATING,TESTING,PREDICTING,TUNING}` ([#4945](https://github.com/PyTorchLightning/pytorch-lightning/pull/4945))

- Added `Trainer.validate()` method to perform one evaluation epoch over the validation set ([#4948](https://github.com/PyTorchLightning/pytorch-lightning/pull/4948))

- Added `LightningEnvironment` for Lightning-specific DDP ([#5915](https://github.com/PyTorchLightning/pytorch-lightning/pull/5915))

- Added `teardown()` hook to LightningDataModule ([#4673](https://github.com/PyTorchLightning/pytorch-lightning/pull/4673))

- Added `auto_insert_metric_name` parameter to `ModelCheckpoint` ([#6277](https://github.com/PyTorchLightning/pytorch-lightning/pull/6277))

- Added arg to `self.log` that enables users to give custom names when dealing with multiple dataloaders ([#6274](https://github.com/PyTorchLightning/pytorch-lightning/pull/6274))

- Added `teardown` method to `BaseProfiler` to enable subclasses defining post-profiling steps outside of `__del__` ([#6370](https://github.com/PyTorchLightning/pytorch-lightning/pull/6370))

- Added `setup` method to `BaseProfiler` to enable subclasses defining pre-profiling steps for every process ([#6633](https://github.com/PyTorchLightning/pytorch-lightning/pull/6633))

- Added no return warning to predict ([#6139](https://github.com/PyTorchLightning/pytorch-lightning/pull/6139))

- Added `Trainer.predict` config validation ([#6543](https://github.com/PyTorchLightning/pytorch-lightning/pull/6543))

- Added `AbstractProfiler` interface ([#6621](https://github.com/PyTorchLightning/pytorch-lightning/pull/6621))

- Added support for including module names for forward in the autograd trace of `PyTorchProfiler` ([#6349](https://github.com/PyTorchLightning/pytorch-lightning/pull/6349))

- Added support for the PyTorch 1.8.1 autograd profiler ([#6618](https://github.com/PyTorchLightning/pytorch-lightning/pull/6618))

- Added `outputs` parameter to callback's `on_validation_epoch_end` & `on_test_epoch_end` hooks ([#6120](https://github.com/PyTorchLightning/pytorch-lightning/pull/6120))

- Added `configure_sharded_model` hook ([#6679](https://github.com/PyTorchLightning/pytorch-lightning/pull/6679))

- Added support for `precision=64`, enabling training with double precision ([#6595](https://github.com/PyTorchLightning/pytorch-lightning/pull/6595))

- Added support for DDP communication hooks ([#6736](https://github.com/PyTorchLightning/pytorch-lightning/issues/6736))

- Added `artifact_location` argument to `MLFlowLogger` which will be passed to the `MlflowClient.create_experiment` call ([#6677](https://github.com/PyTorchLightning/pytorch-lightning/pull/6677))

- Added `model` parameter to precision plugins' `clip_gradients` signature ([#6764](https://github.com/PyTorchLightning/pytorch-lightning/pull/6764))

### Changed

- Renamed `pytorch_lightning.callbacks.swa` to `pytorch_lightning.callbacks.stochastic_weight_avg` ([#6259](https://github.com/PyTorchLightning/pytorch-lightning/pull/6259))

- Refactor `RunningStage` and `TrainerState` usage ([#4945](https://github.com/PyTorchLightning/pytorch-lightning/pull/4945))

- Changed `trainer.evaluating` to return `True` if validating or testing ([#4945](https://github.com/PyTorchLightning/pytorch-lightning/pull/4945))

- Changed `setup()` and `teardown()` stage argument to take any of `{fit,validate,test,predict}` ([#6386](https://github.com/PyTorchLightning/pytorch-lightning/pull/6386))

- Changed profilers to save separate report files per state and rank ([#6621](https://github.com/PyTorchLightning/pytorch-lightning/pull/6621))

- Changed `PyTorchProfiler` to use `torch.autograd.profiler.record_function` to record functions ([#6349](https://github.com/PyTorchLightning/pytorch-lightning/pull/6349))

### Deprecated

- `period` has been deprecated in favor of `every_n_val_epochs` in the `ModelCheckpoint` callback ([#6146](https://github.com/PyTorchLightning/pytorch-lightning/pull/6146))

- Deprecated `trainer.running_sanity_check` in favor of `trainer.sanity_checking` ([#4945](https://github.com/PyTorchLightning/pytorch-lightning/pull/4945))

- Deprecated `Profiler(output_filename)` in favor of `dirpath` and `filename` ([#6621](https://github.com/PyTorchLightning/pytorch-lightning/pull/6621))

- Deprecated `PytorchProfiler(profiled_functions)` in favor of `record_functions` ([#6349](https://github.com/PyTorchLightning/pytorch-lightning/pull/6349))

- Deprecated metrics in favor of `torchmetrics` ([#6505](https://github.com/PyTorchLightning/pytorch-lightning/pull/6505),
    [#6530](https://github.com/PyTorchLightning/pytorch-lightning/pull/6530),
    [#6540](https://github.com/PyTorchLightning/pytorch-lightning/pull/6540),
    [#6547](https://github.com/PyTorchLightning/pytorch-lightning/pull/6547),
    [#6515](https://github.com/PyTorchLightning/pytorch-lightning/pull/6515),
    [#6572](https://github.com/PyTorchLightning/pytorch-lightning/pull/6572),
    [#6573](https://github.com/PyTorchLightning/pytorch-lightning/pull/6573),
    [#6584](https://github.com/PyTorchLightning/pytorch-lightning/pull/6584),
    [#6636](https://github.com/PyTorchLightning/pytorch-lightning/pull/6636),
    [#6637](https://github.com/PyTorchLightning/pytorch-lightning/pull/6637),
    [#6649](https://github.com/PyTorchLightning/pytorch-lightning/pull/6649),
    [#6659](https://github.com/PyTorchLightning/pytorch-lightning/pull/6659),
)

### Removed

- Removed support for passing a bool value to `profiler` argument of Trainer ([#6164](https://github.com/PyTorchLightning/pytorch-lightning/pull/6164))

- Removed no return warning from val/test step ([#6139](https://github.com/PyTorchLightning/pytorch-lightning/pull/6139))

- Removed passing a `ModelCheckpoint` instance to `Trainer(checkpoint_callback)` ([#6166](https://github.com/PyTorchLightning/pytorch-lightning/pull/6166))

- Removed deprecated Trainer argument `enable_pl_optimizer` and `automatic_optimization` ([#6163](https://github.com/PyTorchLightning/pytorch-lightning/pull/6163))

- Removed deprecated metrics ([#6161](https://github.com/PyTorchLightning/pytorch-lightning/pull/6161))
    * from `pytorch_lightning.metrics.functional.classification` removed `to_onehot`, `to_categorical`, `get_num_classes`, `roc`, `multiclass_roc`, `average_precision`, `precision_recall_curve`, `multiclass_precision_recall_curve`
    * from `pytorch_lightning.metrics.functional.reduction` removed `reduce`, `class_reduce`

- Removed deprecated `ModelCheckpoint` arguments `prefix`, `mode="auto"` ([#6162](https://github.com/PyTorchLightning/pytorch-lightning/pull/6162))

- Removed `mode='auto'` from `EarlyStopping` ([#6167](https://github.com/PyTorchLightning/pytorch-lightning/pull/6167))

- Removed legacy references for magic keys in the `Result` object ([#6016](https://github.com/PyTorchLightning/pytorch-lightning/pull/6016))

- Removed deprecated `LightningModule` `hparams` setter ([#6207](https://github.com/PyTorchLightning/pytorch-lightning/pull/6207))

- Removed legacy code to log or include metrics in the progress bar by returning them in a dict with the `"log"/"progress_bar"` magic keys. Use `self.log` instead ([#6734](https://github.com/PyTorchLightning/pytorch-lightning/pull/6734))

- Removed `optimizer_idx` argument from `training_step` in manual optimization ([#6093](https://github.com/PyTorchLightning/pytorch-lightning/pull/6093))

### Fixed

- Set better defaults for `rank_zero_only.rank` when training is launched with SLURM and torchelastic ([#6802](https://github.com/PyTorchLightning/pytorch-lightning/pull/6802/))

- Made the `Plugin.reduce` method more consistent across all Plugins to reflect a mean-reduction by default ([#6011](https://github.com/PyTorchLightning/pytorch-lightning/pull/6011))

- Move lightning module to correct device type when using LightningDistributedWrapper ([#6070](https://github.com/PyTorchLightning/pytorch-lightning/pull/6070))

- Do not print top-k verbose log with `ModelCheckpoint(monitor=None)` ([#6109](https://github.com/PyTorchLightning/pytorch-lightning/pull/6109))

- Fixed csv extension check ([#6436](https://github.com/PyTorchLightning/pytorch-lightning/pull/6436))

- Fixed `ModelCheckpoint(monitor=None, save_last=True)` not saving checkpoints ([#6136](https://github.com/PyTorchLightning/pytorch-lightning/pull/6136))

- Fixed `ModelCheckpoint(save_top_k=0, save_last=True)` not saving the `last` checkpoint ([#6136](https://github.com/PyTorchLightning/pytorch-lightning/pull/6136))

- Fixed `.teardown(stage='fit')` getting called during `trainer.test` ([#6386](https://github.com/PyTorchLightning/pytorch-lightning/pull/6386))

- Fixed `.on_fit_{start,end}()` getting called during `trainer.test` ([#6386](https://github.com/PyTorchLightning/pytorch-lightning/pull/6386))

- Fixed LightningModule `all_gather` on cpu tensors ([#6416](https://github.com/PyTorchLightning/pytorch-lightning/pull/6416))

- Fixed torch distributed not available in setup hook for DDP ([#6506](https://github.com/PyTorchLightning/pytorch-lightning/pull/6506))

- Fixed `EarlyStopping` logic when `min_epochs` or `min_steps` requirement is not met ([#6705](https://github.com/PyTorchLightning/pytorch-lightning/pull/6705))

## [1.2.7] - 2021-04-06

### Fixed

- Fixed resolve a bug with omegaconf and xm.save ([#6741](https://github.com/PyTorchLightning/pytorch-lightning/pull/6741))
- Fixed an issue with IterableDataset when __len__ is not defined ([#6828](https://github.com/PyTorchLightning/pytorch-lightning/pull/6828))
- Sanitize None params during pruning ([#6836](https://github.com/PyTorchLightning/pytorch-lightning/pull/6836))
- Enforce an epoch scheduler interval when using SWA ([#6588](https://github.com/PyTorchLightning/pytorch-lightning/pull/6588))
- Fixed TPU Colab hang issue, post training ([#6816](https://github.com/PyTorchLightning/pytorch-lightning/pull/6816))
- Fixed a bug where `TensorBoardLogger` would give a warning and not log correctly to a symbolic link `save_dir` ([#6730](https://github.com/PyTorchLightning/pytorch-lightning/pull/6730))

## [1.2.6] - 2021-03-30

### Changed

- Changed the behavior of `on_epoch_start` to run at the beginning of validation & test epoch ([#6498](https://github.com/PyTorchLightning/pytorch-lightning/pull/6498))

### Removed

- Removed legacy code to include `step` dictionary returns in `callback_metrics`. Use `self.log_dict` instead. ([#6682](https://github.com/PyTorchLightning/pytorch-lightning/pull/6682))

### Fixed

- Fixed `DummyLogger.log_hyperparams` raising a `TypeError` when running with `fast_dev_run=True` ([#6398](https://github.com/PyTorchLightning/pytorch-lightning/pull/6398))
- Fixed error on TPUs when there was no `ModelCheckpoint` ([#6654](https://github.com/PyTorchLightning/pytorch-lightning/pull/6654))
- Fixed `trainer.test` freeze on TPUs ([#6654](https://github.com/PyTorchLightning/pytorch-lightning/pull/6654))
- Fixed a bug where gradients were disabled after calling `Trainer.predict` ([#6657](https://github.com/PyTorchLightning/pytorch-lightning/pull/6657))
- Fixed bug where no TPUs were detected in a TPU pod env ([#6719](https://github.com/PyTorchLightning/pytorch-lightning/pull/6719))

## [1.2.5] - 2021-03-23

### Changed

- Update Gradient Clipping for the TPU Accelerator ([#6576](https://github.com/PyTorchLightning/pytorch-lightning/pull/6576))
- Refactored setup for typing friendly ([#6590](https://github.com/PyTorchLightning/pytorch-lightning/pull/6590))

### Fixed

- Fixed a bug where `all_gather` would not work correctly with `tpu_cores=8` ([#6587](https://github.com/PyTorchLightning/pytorch-lightning/pull/6587))
- Fixed comparing required versions ([#6434](https://github.com/PyTorchLightning/pytorch-lightning/pull/6434))
- Fixed duplicate logs appearing in console when using the python logging module ([#6275](https://github.com/PyTorchLightning/pytorch-lightning/pull/6275))
- Added Autocast in validation, test and predict modes for Native AMP ([#6565](https://github.com/PyTorchLightning/pytorch-lightning/pull/6565))

Reviewed By: shuyingsunshine21

Differential Revision: D27528929

fbshipit-source-id: 311c88f71461c2c79bbf185e28d7a6d683ccc26f

Summary: `checkpoint_callback` now only accepts boolean values: https://github.com/PyTorchLightning/pytorch-lightning/blob/19e67d18c472c3a03dec4dd9bfcef031e9ca8719/pytorch_lightning/trainer/connectors/callback_connector.py#L65-L73

Reviewed By: shuyingsunshine21

Differential Revision: D27682178

fbshipit-source-id: 9e863aad7a23a76dee8ae5df9f5a78e7a94bfe8a

Summary:
Before: this test would assume only 2 checkpoints were stored: `last.ckpt`, and `FINAL_MODEL_CKPT`
Now: this test asserts that at least these 2 checkpoints are stored. In case the config specifies `save_top_k=-1` for instance, we'd save more checkpoints, causing this test to fail

Since this test is only loading the last and the final outputs, I'm changing the behavior to assert that these checkpoints must be saved and ignoring other checkpoint files that could be generated.

Reviewed By: kazhang

Differential Revision: D27671284

fbshipit-source-id: 0419fb46856d048e7b6eba3ff1dc65b7280a9a90

Summary: Prediction count evaluator needs to gather it's state before computing metrics, otherwise when parallelized across N GPUs, we only get metrics computed from 1/N of the dataset, increasing our eval signal's variance.

Reviewed By: wat3rBro

Differential Revision: D27416864

fbshipit-source-id: b2c5334cd5a38bebcd06c6ace1627a6b71645fdd

Reviewed By: newstzpz

Differential Revision: D27255960

fbshipit-source-id: 1699ff23d2bc610dffc0215a90a7c1c17e3783c3

Summary: Fixing unit test that was not listed due to rebase error.

Reviewed By: newstzpz, wat3rBro

Differential Revision: D27456322

fbshipit-source-id: 519c5c086adfb19104ed99234f4f476eb34a79bc

Summary: Separate unit tests into individual folder based on functionality.

Reviewed By: wat3rBro

Differential Revision: D27132567

fbshipit-source-id: 9a8200be530ca14c7ef42191d59795b05b9800cc

Summary:
Evaluate the predictor generated by previous step.
This diff modify the lightning_train_net to reuse the evaluation logic by adding a `predictor_path` param.
This diff also makes Lightning training backend depends on `cfg.MODEL.DEVICE` so that in evaluate_predictor step, user could set backend by changing model device. This is useful for evaluating int8 quantized model.

Reviewed By: newstzpz

Differential Revision: D27150609

fbshipit-source-id: fb72da3e81db932c0fa479350150720143e09a3e

Summary:
Given that the way to create D2 (https://github.com/facebookresearch/d2go/commit/465cdb842513eb910aa20fcedea1d2edd15dc7b7)go runner and Lightning task are different, get_class was introduced so that in application we could do:
```
if is Lightning:
    task_cls = get_class(classname)
    task = task_cls(cfg)
else:
    runner = create_runner(classname)
```
It turns out that we could need to do that in many places: workflow, binaries.
This diff revert `get_class` and return class in `create_runner` if the class is a Lightning module.

Reviewed By: newstzpz

Differential Revision: D26676595

fbshipit-source-id: c3ce2016d09fe073af4c2dd9f98eea4e59ca621b

Summary: Not d2go.tests is not a library for oss, move utils code to d2go.utils.testing

Reviewed By: zhanghang1989

Differential Revision: D26706933

fbshipit-source-id: 85767b66bbb6c67db05e11823beb4840220b2aa3

Summary:
`checkpoint_callback` is being phased out. Initially, it was a special way to configure checkpoints, but it makes more sense for those callbacks to be included in the general `callbacks` trainer argument. In 1.2.X, `checkpoint_callback` is expected to be a boolean value only.

If `checkpoint_callback=False` **and** an instance of `ModelCheckpoint` is passed in the trainer's `callbacks` arguments, Lightning raises a [misconfiguration error](https://github.com/PyTorchLightning/pytorch-lightning/blob/2f6ce1ae7fff34d16d3707571f6a9a7b0fb0c50a/pytorch_lightning/trainer/connectors/callback_connector.py#L66-L70)

Reviewed By: newstzpz

Differential Revision: D27139315

fbshipit-source-id: 07ad5ea520583a2e46a9cb2a938f98968265c932

Summary: Extend conversion to support ids beyond cocotext format where ids are strings.

Reviewed By: newstzpz

Differential Revision: D27018211

fbshipit-source-id: 7282fd4b9a7e9cd19323235ed1a3c3e7b33cb6b4

Summary:
Pull Request resolved: https://github.com/facebookresearch/d2go/pull/17

Use PyTorch Lightning checkpoint in the test.

Reviewed By: zhanghang1989

Differential Revision: D26962697

fbshipit-source-id: abe635e374c3ada130243f0eaadff34204f04fa1

Summary: migrated to iopath for d2go - rest.

Reviewed By: zhanghang1989

Differential Revision: D26632160

fbshipit-source-id: 182fcd9ad493a483ab7f342f373a7be2de7d10ca

Summary: Pull Request resolved: https://github.com/facebookresearch/d2go/pull/5

Reviewed By: wat3rBro

Differential Revision: D26780956

Pulled By: zhanghang1989

fbshipit-source-id: 26af80bbdf6bcb6af4a8b5d27e655826b34db26a

Summary:
As titled. The OSS version only use PyTorch Lightning while internal version leverages some features(e.g. Manifold integration, every_n_step checkpointing).
This diff splits train_net.main into smaller functions so that they could be shared across OSS and internal versions.

Reviewed By: zhanghang1989

Differential Revision: D26752701

fbshipit-source-id: 7f68e2a81e78193e117517a0ff668ab14b76ea65

Summary: As titled. Make a copy of quantization callback to unblock D2go OSS.

Reviewed By: zhanghang1989

Differential Revision: D26735525

fbshipit-source-id: 12b77f04cfa1361e856b26ea218a262da1fadd88

fbshipit-source-id: f4a8ba78691d8cf46e003ef0bd2e95f170932778