Change default interval to step and add lr in prog_bar in nemo.lightn…

…ing (NVIDIA#10033) Signed-off-by: Hemil Desai <[email protected]> Signed-off-by: Hainan Xu <[email protected]>
hainan-xv · Nov 5, 2024 · bc31182 · bc31182
1 parent 6257a4f
commit bc31182
Show file tree

Hide file tree

Showing 2 changed files with 19 additions and 19 deletions.
diff --git a/nemo/lightning/pytorch/optim/base.py b/nemo/lightning/pytorch/optim/base.py
@@ -152,7 +152,7 @@ def optimizers(self, model) -> List[Optimizer]:
     def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx) -> None:
         if self._optimizers is not None:
             lr = self._optimizers[0].param_groups[0]['lr']
-            pl_module.log('lr', lr, rank_zero_only=True, batch_size=1)
+            pl_module.log('lr', lr, rank_zero_only=True, batch_size=1, prog_bar=True)
 
     def __call__(self, model: L.LightningModule, megatron_parallel=None) -> OptimizerLRScheduler:
         """Calls the setup and optimizers methods.

diff --git a/nemo/lightning/pytorch/optim/lr_scheduler.py b/nemo/lightning/pytorch/optim/lr_scheduler.py
@@ -25,7 +25,7 @@ def __init__(
         warmup_ratio: Optional[float] = None,
         max_steps: int = 10,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -68,7 +68,7 @@ def __init__(
         hold_ratio: Optional[float] = None,
         max_steps: int = 10,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -111,7 +111,7 @@ def __init__(
         self,
         max_steps: int = 10,
         min_lr: float = 1e-5,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -142,7 +142,7 @@ def __init__(
         self,
         max_steps: int = 10,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -176,7 +176,7 @@ def __init__(
         warmup_ratio: Optional[float] = None,
         max_steps: int = 10,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -218,7 +218,7 @@ def __init__(
         max_steps: int = 10,
         decay_rate: float = 0.5,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -252,7 +252,7 @@ def __init__(
         self,
         max_steps: int = 10,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -283,7 +283,7 @@ def __init__(
         self,
         max_steps: int = 10,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -314,7 +314,7 @@ def __init__(
         self,
         max_steps: int = 10,
         min_lr: float = 0.0,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -347,7 +347,7 @@ def __init__(
         min_lr: float = 0.0,
         power: float = 1.0,
         cycle: bool = False,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -384,7 +384,7 @@ def __init__(
         min_lr: float = 0.0,
         power: float = 1.0,
         cycle: bool = False,
-        interval: str = "epoch",
+        interval: str = "step",
         frequency: int = 1,
         monitor: str = "val_loss",
     ):
@@ -415,13 +415,13 @@ def scheduler(self, model, optimizer):
 class CosineAnnealingScheduler(LRSchedulerModule):
     def __init__(
         self,
-        max_steps=10,
-        warmup_steps=750,
-        constant_steps=80000,
-        min_lr=int(6e-5),
-        interval="epoch",
-        frequency=1,
-        monitor="val_loss",
+        max_steps: int = 10,
+        warmup_steps: int = 750,
+        constant_steps: int = 80000,
+        min_lr: float = 6e-5,
+        interval: str = "step",
+        frequency: int = 1,
+        monitor: str = "val_loss",
     ):
         super().__init__()
         self.max_steps = max_steps