Merge branch 'master' into metrics/fbeta

CHANGELOG.md

@@ -52,6 +52,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
+- Consistently use `step=trainer.global_step` in `LearningRateMonitor` independently of `logging_interval` ([#4376](https://github.com/PyTorchLightning/pytorch-lightning/pull/4376))
+
+
 - Tuner algorithms will be skipped if `fast_dev_run=True` ([#3903](https://github.com/PyTorchLightning/pytorch-lightning/pull/3903))
 
 

pytorch_lightning/callbacks/lr_monitor.py

@@ -114,7 +114,7 @@ def on_train_epoch_start(self, trainer, *args, **kwargs):
             latest_stat = self._extract_stats(trainer, interval)
 
             if trainer.logger is not None and latest_stat:
-                trainer.logger.log_metrics(latest_stat, step=trainer.current_epoch)
+                trainer.logger.log_metrics(latest_stat, step=trainer.global_step)
 
     def _extract_stats(self, trainer, interval: str) -> Dict[str, float]:
         latest_stat = {}

pytorch_lightning/tuner/batch_size_scaling.py

@@ -102,7 +102,7 @@ def scale_batch_size(trainer,
         trainer.progress_bar_callback.disable()
 
     # Initially we just double in size until an OOM is encountered
-    new_size = _adjust_batch_size(trainer, value=init_val)  # initially set to init_val
+    new_size = _adjust_batch_size(trainer, batch_arg_name, value=init_val)  # initially set to init_val
     if mode == 'power':
         new_size = _run_power_scaling(trainer, model, new_size, batch_arg_name, max_trials, **fit_kwargs)
     elif mode == 'binsearch':
@@ -231,7 +231,7 @@ def _run_binsearch_scaling(trainer, model, new_size, batch_arg_name, max_trials,
                 garbage_collection_cuda()
                 high = new_size
                 midval = (high + low) // 2
-                new_size, _ = _adjust_batch_size(trainer, value=midval, desc='failed')
+                new_size, _ = _adjust_batch_size(trainer, batch_arg_name, value=midval, desc='failed')
                 if high - low <= 1:
                     break
             else: