Lightning-AI/pytorch-lightning 0.10.0

Buffer release before 1.0

on GitHub

This release is a buffer in case 1.0 breaks any compatibility for people who upgrade. 0.10.0 has all the bug fixes and features of 1.0 but is 100% backward compatible. The 1.0 release following in the next 24 hours.

Overview

The major changes are:

• Results objects are deprecated (we hated them too haha)
• This means dataflow and logging have been decoupled

To log:

def any_step(...):
   self.log('something', i_computed)

Separately, return whatever you want from methods:

def training_step(...):
  return loss

or

def training_step(...):
   return {'loss': loss, 'whatever': [1, 'want']}

Detail changes

Added

• Added new Metrics API. (#3868, [#3921)
• Enable PyTorch 1.7 compatibility (#3541)
• Added LightningModule.to_torchscript to support exporting as ScriptModule (#3258)
• Added warning when dropping unpicklable hparams (#2874)
• Added EMB similarity (#3349)
• Added ModelCheckpoint.to_yaml method (#3048)
• Allow ModelCheckpoint monitor to be None, meaning it will always save ([3630)
• Disabled optimizers setup during testing (#3059)
• Added support for datamodules to save and load checkpoints when training (#3563
• Added support for datamodule in learning rate finder (#3425)
• Added gradient clip test for native AMP (#3754)
• Added dist lib to enable syncing anything across devices (#3762)
• Added broadcast to TPUBackend (#3814)
• Added XLADeviceUtils class to check XLA device type (#3274)

Changed

• Refactored accelerator backends:
  • moved TPU xxx_step to backend (#3118)
  • refactored DDP backend forward (#3119)
  • refactored GPU backend __step (#3120)
  • refactored Horovod backend (#3121, #3122)
  • remove obscure forward call in eval + CPU backend ___step (#3123)
  • reduced all simplified forward (#3126)
  • added hook base method (#3127)
  • refactor eval loop to use hooks - use test_mode for if so we can split later (#3129)
  • moved ___step_end hooks (#3130)
  • training forward refactor (#3134)
  • training AMP scaling refactor (#3135)
  • eval step scaling factor (#3136)
  • add eval loop object to streamline eval loop (#3138)
  • refactored dataloader process hook (#3139)
  • refactored inner eval loop (#3141)
  • final inner eval loop hooks (#3154)
  • clean up hooks in run_evaluation (#3156)
  • clean up data reset (#3161)
  • expand eval loop out (#3165)
  • moved hooks around in eval loop (#3195)
  • remove _evaluate fx (#3197)
  • Trainer.fit hook clean up (#3198)
  • DDPs train hooks (#3203)
  • refactor DDP backend (#3204, #3207, #3208, #3209, #3210)
  • reduced accelerator selection (#3211)
  • group prepare data hook (#3212)
  • added data connector (#3285)
  • modular is_overridden (#3290)
  • adding Trainer.tune() (#3293)
  • move run_pretrain_routine -> setup_training (#3294)
  • move train outside of setup training (#3297)
  • move prepare_data to data connector (#3307)
  • moved accelerator router (#3309)
  • train loop refactor - moving train loop to own object (#3310, #3312, #3313, #3314)
  • duplicate data interface definition up into DataHooks class (#3344)
  • inner train loop (#3359, #3361, #3362, #3363, #3365, #3366, #3367, #3368, #3369, #3370, #3371, #3372, #3373, #3374, #3375, #3376, #3385, #3388, #3397)
  • all logging related calls in a connector (#3395)
  • device parser (#3400, #3405)
  • added model connector (#3407)
  • moved eval loop logging to loggers (#3408)
  • moved eval loop (#3412[#3408)
  • trainer/separate argparse (#3421, #3428, #3432)
  • move lr_finder (#3434)
  • organize args (##3435, #3442, #3447, #3448, #3449, #3456)
  • move specific accelerator code (#3457)
  • group connectors (#3472)
  • accelerator connector methods x/n (#3469, #3470, #3474)
  • merge backends (#3476, #3477, #3478, #3480, #3482)
  • apex plugin (#3502)
  • precision plugins (#3504)
  • Result - make monitor default to checkpoint_on to simplify (#3571)
  • reference to the Trainer on the LightningDataModule (#3684)
  • add .log to lightning module (#3686, #3699, #3701, #3704, #3715)
  • enable tracking original metric when step and epoch are both true (#3685)
  • deprecated results obj, added support for simpler comms (#3681)
  • move backends back to individual files (#3712)
  • fixes logging for eval steps (#3763)
  • decoupled DDP, DDP spawn (#3733, #3766, #3767, #3774, #3802, #3806)
  • remove weight loading hack for ddp_cpu (#3808)
  • separate torchelastic from DDP (#3810)
  • separate SLURM from DDP (#3809)
  • decoupled DDP2 (#3816)
  • bug fix with logging val epoch end + monitor (#3812)
  • decoupled DDP, DDP spawn (#3733, #3817, #3819, #3927)
  • callback system and init DDP (#3836)
  • adding compute environments (#3837, [#3842)
  • epoch can now log independently (#3843)
  • test selecting the correct backend. temp backends while slurm and TorchElastic are decoupled (#3848)
  • fixed init_slurm_connection causing hostname errors (#3856)
  • moves init apex from LM to apex connector (#3923)
  • moves sync bn to each backend (#3925)
  • moves configure ddp to each backend (#3924)
• Deprecation warning (#3844)
• Changed LearningRateLogger to LearningRateMonitor (#3251)
• Used fsspec instead of gfile for all IO (#3320)
  • Swaped torch.load for fsspec load in DDP spawn backend (#3787)
  • Swaped torch.load for fsspec load in cloud_io loading (#3692)
  • Added support for to_disk() to use remote filepaths with fsspec (#3930)
  • Updated model_checkpoint's to_yaml to use fsspec open (#3801)
  • Fixed fsspec is inconsistant when doing fs.ls (#3805)
• Refactor GPUStatsMonitor to improve training speed (#3257)
• Changed IoU score behavior for classes absent in target and pred (#3098)
• Changed IoU remove_bg bool to ignore_index optional int (#3098)
• Changed defaults of save_top_k and save_last to None in ModelCheckpoint (#3680)
• row_log_interval and log_save_interval are now based on training loop's global_step instead of epoch-internal batch index (#3667)
• Silenced some warnings. verified ddp refactors (#3483)
• Cleaning up stale logger tests (#3490)
• Allow ModelCheckpoint monitor to be None (#3633)
• Enable None model checkpoint default (#3669)
• Skipped best_model_path if checkpoint_callback is None (#2962)
• Used raise .. from .. to explicitly chain exceptions (#3750)
• Mocking loggers (#3596, #3617, #3851, #3859, #3884, #3853, #3910, #3889, #3926)
• Write predictions in LightningModule instead of EvalResult [#3882

Deprecated

• Deprecated TrainResult and EvalResult, use self.log and self.write from the LightningModule to log metrics and write predictions. training_step can now only return a scalar (for the loss) or a dictionary with anything you want. (#3681)
• Deprecate early_stop_callback Trainer argument (#3845)
• Rename Trainer arguments row_log_interval >> log_every_n_steps and log_save_interval >> flush_logs_every_n_steps (#3748)

Removed

• Removed experimental Metric API (#3868, #3943, #3949, #3946), listed changes before final removal:
  • Added EmbeddingSimilarity metric (#3349, [#3358)
  • Added hooks to metric module interface (#2528)
  • Added error when AUROC metric is used for multiclass problems (#3350)
  • Fixed ModelCheckpoint with save_top_k=-1 option not tracking the best models when a monitor metric is available (#3735)
  • Fixed counter-intuitive error being thrown in Accuracy metric for zero target tensor (#3764)
  • Fixed aggregation of metrics (#3517)
  • Fixed Metric aggregation (#3321)
  • Fixed RMSLE metric (#3188)
  • Renamed reduction to class_reduction in classification metrics (#3322)
  • Changed class_reduction similar to sklearn for classification metrics (#3322)
  • Renaming of precision recall metric (#3308)

Fixed

• Fixed on_train_batch_start hook to end epoch early (#3700)
• Fixed num_sanity_val_steps is clipped to limit_val_batches (#2917)
• Fixed ONNX model save on GPU (#3145)
• Fixed GpuUsageLogger to work on different platforms (#3008)
• Fixed auto-scale batch size not dumping auto_lr_find parameter (#3151)
• Fixed batch_outputs with optimizer frequencies (#3229)
• Fixed setting batch size in LightningModule.datamodule when using auto_scale_batch_size (#3266)
• Fixed Horovod distributed backend compatibility with native AMP (#3404)
• Fixed batch size auto scaling exceeding the size of the dataset (#3271)
• Fixed getting experiment_id from MLFlow only once instead of each training loop (#3394)
• Fixed overfit_batches which now correctly disables shuffling for the training loader. (#3501)
• Fixed gradient norm tracking for row_log_interval > 1 (#3489)
• Fixed ModelCheckpoint name formatting ([3164)
• Fixed auto-scale batch size (#3151)
• Fixed example implementation of AutoEncoder (#3190)
• Fixed invalid paths when remote logging with TensorBoard (#3236)
• Fixed change t() to transpose() as XLA devices do not support .t() on 1-dim tensor (#3252)
• Fixed (weights only) checkpoints loading without PL (#3287)
• Fixed gather_all_tensors cross GPUs in DDP (#3319)
• Fixed CometML save dir (#3419)
• Fixed forward key metrics (#3467)
• Fixed normalize mode at confusion matrix (replace NaNs with zeros) (#3465)
• Fixed global step increment in training loop when training_epoch_end hook is used (#3673)
• Fixed dataloader shuffling not getting turned off with overfit_batches > 0 and distributed_backend = "ddp" (#3534)
• Fixed determinism in DDPSpawnBackend when using seed_everything in main process (#3335)
• Fixed ModelCheckpoint period to actually save every period epochs (#3630)
• Fixed val_progress_bar total with num_sanity_val_steps (#3751)
• Fixed Tuner dump: add current_epoch to dumped_params (#3261)
• Fixed current_epoch and global_step properties mismatch between Trainer and LightningModule (#3785)
• Fixed learning rate scheduler for optimizers with internal state (#3897)
• Fixed tbptt_reduce_fx when non-floating tensors are logged (#3796)
• Fixed model checkpoint frequency (#3852)
• Fixed logging non-tensor scalar with result breaks subsequent epoch aggregation (#3855)
• Fixed TrainerEvaluationLoopMixin activates model.train() at the end (#3858)
• Fixed overfit_batches when using with multiple val/test_dataloaders (#3857)
• Fixed enables training_step to return None (#3862)
• Fixed init nan for checkpointing (#3863)
• Fixed for load_from_checkpoint (#2776)
• Fixes incorrect batch_sizes when Dataloader returns a dict with multiple tensors (#3668)
• Fixed unexpected signature for validation_step (#3947)

Contributors

@abrahambotros, @akihironitta, @ananthsub, @ananyahjha93, @awaelchli, @Borda, @c00k1ez, @carmocca, @f4hy, @GimmickNG, @jbschiratti, @justusschock, @LeeJZh, @lezwon, @Lucas-Steinmann, @maxjeblick, @monney, @mpariente, @nateraw, @nrupatunga, @patrickorlando, @PhilJd, @rohitgr7, @s-rog, @ShomyLiu, @SkafteNicki, @Sordie, @teddykoker, @tgaddair, @Vozf, @williamFalcon, @XDynames, @ydcjeff

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