few fixes

InnerEye/ML/SSL/lightning_modules/ssl_classifier_module.py

@@ -86,7 +86,7 @@ def training_step(self, batch: Any, batch_id: int, *args: Any, **kwargs: Any) ->
 
  def validation_step(self, batch: Any, batch_id: int, *args: Any, **kwargs: Any) -> None: # type: ignore
  loss = self.shared_step(batch, is_training=False)
- self.log('val/loss', loss, on_step=False, on_epoch=True, sync_dist=False)
+ self.log('val/loss', loss, on_step=False, on_epoch=True, sync_dist=True)
  for metric in self.val_metrics:
  self.log(f"val/{metric.name}", metric, on_epoch=True, on_step=False)
 

InnerEye/ML/lightning_base.py

@@ -383,7 +383,7 @@ def log_on_epoch(self,
  if isinstance(value, numbers.Number):
  value = torch.tensor(value, dtype=torch.float, device=self.device)
  prefix = TRAIN_PREFIX if is_training else VALIDATION_PREFIX
- sync_dist = True if sync_dist_override is None else sync_dist_override
+ sync_dist = self.use_sync_dist if sync_dist_override is None else sync_dist_override
  self.log(prefix + metric_name, value,
  sync_dist=sync_dist,
  on_step=False, on_epoch=True,

InnerEye/ML/lightning_models.py

@@ -143,8 +143,8 @@ def compute_metrics(self, cropped_sample: CroppedSample, segmentation: torch.Ten
  self.log_on_epoch(name=MetricType.SUBJECT_COUNT,
  value=num_subjects,
  is_training=is_training,
- reduce_fx=sum,
- sync_dist_op=None)
+ reduce_fx=torch.sum,
+ sync_dist_op="sum")
 
  def training_or_validation_epoch_end(self, is_training: bool) -> None:
  """