Lightning-AI/pytorch-lightning 1.3.0

Lightning CLI, PyTorch Profiler, Improved Early Stopping

on GitHub

Today we are excited to announce Lightning 1.3, containing highly anticipated new features including a new Lightning CLI, improved TPU support, integrations such as PyTorch profiler, new early stopping strategies, predict and validate trainer routines, and more.

https://medium.com/pytorch/pytorch-lightning-1-3-lightning-cli-pytorch-profiler-improved-early-stopping-6e0ffd8deb29

[1.3.0] - 2021-05-06

Added

• Added support for the EarlyStopping callback to run at the end of the training epoch (#6944)
• Added synchronization points before and after setup hooks are run (#7202)
• Added a teardown hook to ClusterEnvironment (#6942)
• Added utils for metrics to scalar conversions (#7180)
• Added utils for NaN/Inf detection for gradients and parameters (#6834)
• Added more explicit exception message when trying to execute trainer.test() or trainer.validate() with fast_dev_run=True (#6667)
• Added LightningCLI class to provide simple reproducibility with minimum boilerplate training CLI (#4492, #6862, #7156, #7299)
• Added gradient_clip_algorithm argument to Trainer for gradient clipping by value (#6123).
• Added a way to print to terminal without breaking up the progress bar (#5470)
• Added support to checkpoint after training steps in ModelCheckpoint callback (#6146)
• Added TrainerStatus.{INITIALIZING,RUNNING,FINISHED,INTERRUPTED} (#7173)
• Added Trainer.validate() method to perform one evaluation epoch over the validation set (#4948)
• Added LightningEnvironment for Lightning-specific DDP (#5915)
• Added teardown() hook to LightningDataModule (#4673)
• Added auto_insert_metric_name parameter to ModelCheckpoint (#6277)
• Added arg to self.log that enables users to give custom names when dealing with multiple dataloaders (#6274)
• Added teardown method to BaseProfiler to enable subclasses defining post-profiling steps outside of __del__ (#6370)
• Added setup method to BaseProfiler to enable subclasses defining pre-profiling steps for every process (#6633)
• Added no return warning to predict (#6139)
• Added Trainer.predict config validation (#6543)
• Added AbstractProfiler interface (#6621)
• Added support for including module names for forward in the autograd trace of PyTorchProfiler (#6349)
• Added support for the PyTorch 1.8.1 autograd profiler (#6618)
• Added outputs parameter to callback's on_validation_epoch_end & on_test_epoch_end hooks (#6120)
• Added configure_sharded_model hook (#6679)
• Added support for precision=64, enabling training with double precision (#6595)
• Added support for DDP communication hooks (#6736)
• Added artifact_location argument to MLFlowLogger which will be passed to the MlflowClient.create_experiment call (#6677)
• Added model parameter to precision plugins' clip_gradients signature (#6764, #7231)
• Added is_last_batch attribute to Trainer (#6825)
• Added LightningModule.lr_schedulers() for manual optimization (#6567)
• Added MpModelWrapper in TPU Spawn (#7045)
• Added max_time Trainer argument to limit training time (#6823)
• Added on_predict_{batch,epoch}_{start,end} hooks (#7141)
• Added new EarlyStopping parameters stopping_threshold and divergence_threshold (#6868)
• Added debug flag to TPU Training Plugins (PT_XLA_DEBUG) (#7219)
• Added new UnrepeatedDistributedSampler and IndexBatchSamplerWrapper for tracking distributed predictions (#7215)
• Added trainer.predict(return_predictions=None|False|True) (#7215)
• Added BasePredictionWriter callback to implement prediction saving (#7127)
• Added trainer.tune(scale_batch_size_kwargs, lr_find_kwargs) arguments to configure the tuning algorithms (#7258)
• Added tpu_distributed check for TPU Spawn barrier (#7241)
• Added device updates to TPU Spawn for Pod training (#7243)
• Added warning when missing Callback and using resume_from_checkpoint (#7254)
• DeepSpeed single file saving (#6900)
• Added Training type Plugins Registry (#6982, #7063, #7214, #7224)
• Add ignore param to save_hyperparameters (#6056)

Changed

• Changed LightningModule.truncated_bptt_steps to be property (#7323)
• Changed EarlyStopping callback from by default running EarlyStopping.on_validation_end if only training is run. Set check_on_train_epoch_end to run the callback at the end of the train epoch instead of at the end of the validation epoch (#7069)
• Renamed pytorch_lightning.callbacks.swa to pytorch_lightning.callbacks.stochastic_weight_avg (#6259)
• Refactor RunningStage and TrainerState usage (#4945, #7173)
  • Added RunningStage.SANITY_CHECKING
  • Added TrainerFn.{FITTING,VALIDATING,TESTING,PREDICTING,TUNING}
  • Changed trainer.evaluating to return True if validating or testing
• Changed setup() and teardown() stage argument to take any of {fit,validate,test,predict} (#6386)
• Changed profilers to save separate report files per state and rank (#6621)
• The trainer no longer tries to save a checkpoint on exception or run callback's on_train_end functions (#6864)
• Changed PyTorchProfiler to use torch.autograd.profiler.record_function to record functions (#6349)
• Disabled lr_scheduler.step() in manual optimization (#6825)
• Changed warnings and recommendations for dataloaders in ddp_spawn (#6762)
• pl.seed_everything will now also set the seed on the DistributedSampler (#7024)
• Changed default setting for communication of multi-node training using DDPShardedPlugin (#6937)
• trainer.tune() now returns the tuning result (#7258)
• LightningModule.from_datasets() now accepts IterableDataset instances as training datasets. (#7503)
• Changed resume_from_checkpoint warning to an error when the checkpoint file does not exist (#7075)
• Automatically set sync_batchnorm for training_type_plugin (#6536)
• Allowed training type plugin to delay optimizer creation (#6331)
• Removed ModelSummary validation from train loop on_trainer_init (#6610)
• Moved save_function to accelerator (#6689)
• Updated DeepSpeed ZeRO (#6546, #6752, #6142, #6321)
• Improved verbose logging for EarlyStopping callback (#6811)
• Run ddp_spawn dataloader checks on Windows (#6930)
• Updated mlflow with using resolve_tags (#6746)
• Moved save_hyperparameters to its own function (#7119)
• Replaced _DataModuleWrapper with __new__ (#7289)
• Reset current_fx properties on lightning module in teardown (#7247)
• Auto-set DataLoader.worker_init_fn with seed_everything (#6960)
• Remove model.trainer call inside of dataloading mixin (#7317)
• Split profilers module (#6261)
• Ensure accelerator is valid if running interactively (#5970)
• Disabled batch transfer in DP mode (#6098)

Deprecated

• Deprecated outputs in both LightningModule.on_train_epoch_end and Callback.on_train_epoch_end hooks (#7339)
• Deprecated Trainer.truncated_bptt_steps in favor of LightningModule.truncated_bptt_steps (#7323)
• Deprecated outputs in both LightningModule.on_train_epoch_end and Callback.on_train_epoch_end hooks (#7339)
• Deprecated LightningModule.grad_norm in favor of pytorch_lightning.utilities.grads.grad_norm (#7292)
• Deprecated the save_function property from the ModelCheckpoint callback (#7201)
• Deprecated LightningModule.write_predictions and LightningModule.write_predictions_dict (#7066)
• Deprecated TrainerLoggingMixin in favor of a separate utilities module for metric handling (#7180)
• Deprecated TrainerTrainingTricksMixin in favor of a separate utilities module for NaN/Inf detection for gradients and parameters (#6834)
• period has been deprecated in favor of every_n_val_epochs in the ModelCheckpoint callback (#6146)
• Deprecated trainer.running_sanity_check in favor of trainer.sanity_checking (#4945)
• Deprecated Profiler(output_filename) in favor of dirpath and filename (#6621)
• Deprecated PytorchProfiler(profiled_functions) in favor of record_functions (#6349)
• Deprecated @auto_move_data in favor of trainer.predict (#6993)
• Deprecated Callback.on_load_checkpoint(checkpoint) in favor of Callback.on_load_checkpoint(trainer, pl_module, checkpoint) (#7253)
• Deprecated metrics in favor of torchmetrics (#6505, #6530, #6540, #6547, #6515, #6572, #6573, #6584, #6636, #6637, #6649, #6659, #7131)
• Deprecated the LightningModule.datamodule getter and setter methods; access them through Trainer.datamodule instead (#7168)
• Deprecated the use of Trainer(gpus="i") (string) for selecting the i-th GPU; from v1.5 this will set the number of GPUs instead of the index (#6388)

Removed

• Removed the exp_save_path property from the LightningModule (#7266)
• Removed training loop explicitly calling EarlyStopping.on_validation_end if no validation is run (#7069)
• Removed automatic_optimization as a property from the training loop in favor of LightningModule.automatic_optimization (#7130)
• Removed evaluation loop legacy returns for *_epoch_end hooks (#6973)
• Removed support for passing a bool value to profiler argument of Trainer (#6164)
• Removed no return warning from val/test step (#6139)
• Removed passing a ModelCheckpoint instance to Trainer(checkpoint_callback) (#6166)
• Removed deprecated Trainer argument enable_pl_optimizer and automatic_optimization (#6163)
• Removed deprecated metrics (#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)
• Removed mode='auto' from EarlyStopping (#6167)
• Removed epoch and step arguments from ModelCheckpoint.format_checkpoint_name(), these are now included in the metrics argument (#7344)
• Removed legacy references for magic keys in the Result object (#6016)
• Removed deprecated LightningModule hparams setter (#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)
• Removed trainer.fit() return value of 1. It has no return now (#7237)
• Removed logger_connector legacy code (#6733)
• Removed unused mixin attributes (#6487)

Fixed

• Fixed NaN errors in progress bars when training with iterable datasets with no length defined (#7306)
• Fixed attaching train and validation dataloaders when reload_dataloaders_every_epoch=True and num_sanity_val_steps=0 (#7207)
• Added a barrier in the accelerator teardown to synchronize processes before execution finishes (#6814)
• Fixed multi-node DDP sub-process launch by using local_rank instead of global_rank for main process assertion (#7061)
• Fixed incorrect removal of WORLD_SIZE environment variable in DDP training when launching with torch distributed/torchelastic (#6942)
• Made the Plugin.reduce method more consistent across all Plugins to reflect a mean-reduction by default (#6011)
• Move lightning module to correct device type when using LightningDistributedWrapper (#6070)
• Do not print top-k verbose log with ModelCheckpoint(monitor=None) (#6109)
• Fixed ModelCheckpoint(save_top_k=0, save_last=True) not saving the last checkpoint (#6136)
• Fixed .teardown(stage='fit') and .on_fit_{start,end}() getting called during trainer.test (#6386)
• Fixed LightningModule all_gather on cpu tensors (#6416)
• Fixed torch distributed not available in setup hook for DDP (#6506)
• Fixed trainer.tuner.{lr_find,scale_batch_size} not setting the Trainer state properly (#7258)
• Fixed bug where the learning rate schedulers did not follow the optimizer frequencies (#4868)
• Fixed pickle error checker to now check for pickle.PickleError to catch all pickle errors (#6917)
• Fixed a bug where the outputs object passed to LightningModule.training_epoch_end was different from the object passed to the on_train_end_epoch hook (#6969)
• Fixed a bug where the outputs passed to train_batch_end would be listed even when using a single optimizer and no truncated backprop through time steps (#6969)
• Fixed bug for trainer error handling which would cause hang for distributed training (#6864)
• Fixed self.device not returning the correct device in replicas of data-parallel (#6414)
• Fixed lr_find trying beyond num_training steps and suggesting a too high learning rate (#7076)
• Fixed logger creating incorrect version folder in DDP with repeated Trainer.fit calls (#7077)
• Fixed metric objects passed directly to self.log not being reset correctly (#7055)
• Fixed CombinedLoader in distributed settings for validation / testing (#7102)
• Fixed the save_dir in WandbLogger when the run was initiated externally (#7106)
• Fixed num_sanity_val_steps affecting reproducibility of training data shuffling (#7014)
• Fixed resetting device after fitting/evaluating/predicting (#7188)
• Fixed bug where trainer.tuner.scale_batch_size(max_trials=0) would not return the correct batch size result (#7262)
• Fixed metrics not being properly logged with precision=16 and manual_optimization (#7228)
• Fixed BaseFinetuning properly reloading optimizer_states when using resume_from_checkpoint (#6891)
• Fixed parameters_to_ignore not properly set to DDPWrapper (#7239)
• Fixed parsing of fast_dev_run=True with the built-in ArgumentParser (#7240)
• Fixed handling an IterableDataset that fails to produce a batch at the beginning of an epoch (#7294)
• Fixed LightningModule.save_hyperparameters() when attempting to save an empty container (#7268)
• Fixed apex not properly instantiated when running with ddp (#7274)
• Fixed optimizer state not moved to GPU (#7277)
• Fixed custom init args for WandbLogger (#6989)
• Fixed a bug where an error would be raised if the train dataloader sometimes produced None for a batch (#7342)
• Fixed examples (#6600, #6638, #7096, #7246, #6357, #6476, #6294, #6373, #6088, #7398)
• Resolved schedule step bug for PyTorch Profiler (#6674, #6681)
• Updated logic for checking TPUs availability (#6767)
• Resolve TPU miss rendezvous (#6781)
• Fixed auto-scaling mode when calling tune method on trainer (#7321)
• Fixed finetuning complex models correctly unfreezes (#6880)
• Ensure we set the eval/train flag correctly on accelerator model (#6877)
• Set better defaults for rank_zero_only.rank when training is launched with SLURM and torchelastic (#6802)
• Fixed matching the number of outputs of backward with forward for AllGatherGrad (#6625)
• Fixed the gradient_clip_algorithm has no effect (#6928)
• Fixed CUDA OOM detection and handling (#6934)
• Fixed unfreeze_and_add_param_group expects modules rather than module (#6822)
• Fixed DPP + SyncBN when move on device (#6838)
• Fixed missing arguments in lr_find call (#6784)
• Fixed set_default_tensor_type to torch.DoubleTensor with precision=64 (#7108)
• Fixed NeptuneLogger.log_text(step=None) (#7194)
• Fixed importing torchtext batch (#6365, #6323, #6211)

Contributors

@akihironitta, @alessiobonfiglio, @amisev, @amogkam, @ananthsub, @ArvinZhuang, @ashleve, @asnorkin, @awaelchli, @BloodAxe, @bmahlbrand, @Borda, @borisdayma, @camruta, @carmocca, @ceshine, @dbonner, @dhkim0225, @EdwardJB, @EliaCereda, @EricCousineau-TRI, @ethanwharris, @FlorianMF, @hemildesai, @ifsheldon, @kaushikb11, @mauvilsa, @maxfrei750, @mesejo, @ramonemiliani93, @rohitgr7, @s-rog, @sadiqj, @scart97, @SeanNaren, @shuyingsunshine21, @SkafteNicki, @SpontaneousDuck, @stllfe, @tchaton, @THasthika, @vballoli

If we forgot someone due to not matching commit email with GitHub account, let us know :]