Merge release 2.0.2 #573

Author: ashleve

configs/trainer/ddp.yaml

@@ -1,11 +1,7 @@
 defaults:
   - default.yaml
 
-# use "ddp_spawn" instead of "ddp",
-# it's slower but normal "ddp" currently doesn't work ideally with hydra
-# https://github.com/facebookresearch/hydra/issues/2070
-# https://pytorch-lightning.readthedocs.io/en/latest/accelerators/gpu_intermediate.html#distributed-data-parallel-spawn
-strategy: ddp_spawn
+strategy: ddp
 
 accelerator: gpu
 devices: 4

src/models/mnist_module.py

@@ -97,7 +97,7 @@ def on_validation_epoch_end(self):
         self.val_acc_best(acc)  # update best so far val acc
         # log `val_acc_best` as a value through `.compute()` method, instead of as a metric object
         # otherwise metric would be reset by lightning after each epoch
-        self.log("val/acc_best", self.val_acc_best.compute(), prog_bar=True)
+        self.log("val/acc_best", self.val_acc_best.compute(), sync_dist=True, prog_bar=True)
 
     def test_step(self, batch: Any, batch_idx: int):
         loss, preds, targets = self.model_step(batch)