initial commit:

- added val-interval argument. Eval and checkpointing is only applied every val-interval epochs.
- Changed `float` to `Optional[float]` in typing Scheduler step function parameter `metric`
- Skipping step of base scheduler in plateau scheduler to avoid `TypeError` when converting `None` to `float`

Author: Tony

timm/scheduler/plateau_lr.py

@@ -5,7 +5,7 @@
 Hacked together by / Copyright 2020 Ross Wightman
 """
 import torch
-from typing import List
+from typing import List, Optional
 
 from .scheduler import Scheduler
 
@@ -86,12 +86,14 @@ def step(self, epoch, metric=None):
                     param_group['lr'] = self.restore_lr[i]
                 self.restore_lr = None
 
-            self.lr_scheduler.step(metric, epoch)  # step the base scheduler
+            # step the base scheduler if metric given
+            if metric is not None:
+                self.lr_scheduler.step(metric, epoch)
 
             if self._is_apply_noise(epoch):
                 self._apply_noise(epoch)
 
-    def step_update(self, num_updates: int, metric: float = None):
+    def step_update(self, num_updates: int, metric: Optional[float] = None):
         return None
 
     def _apply_noise(self, epoch):

timm/scheduler/scheduler.py

@@ -74,14 +74,14 @@ def _get_values(self, t: int, on_epoch: bool = True) -> Optional[List[float]]:
             return None
         return self._get_lr(t)
 
-    def step(self, epoch: int, metric: float = None) -> None:
+    def step(self, epoch: int, metric: Optional[float] = None) -> None:
         self.metric = metric
         values = self._get_values(epoch, on_epoch=True)
         if values is not None:
             values = self._add_noise(values, epoch)
             self.update_groups(values)
 
-    def step_update(self, num_updates: int, metric: float = None):
+    def step_update(self, num_updates: int, metric: Optional[float] = None):
         self.metric = metric
         values = self._get_values(num_updates, on_epoch=False)
         if values is not None:

train.py

@@ -370,6 +370,8 @@
                    help='worker seed mode (default: all)')
 group.add_argument('--log-interval', type=int, default=50, metavar='N',
                    help='how many batches to wait before logging training status')
+group.add_argument('--val-interval', type=int, default=1, metavar='N',
+                   help='how many epochs between validation and checkpointing')
 group.add_argument('--recovery-interval', type=int, default=0, metavar='N',
                    help='how many batches to wait before writing recovery checkpoint')
 group.add_argument('--checkpoint-hist', type=int, default=10, metavar='N',
@@ -1013,6 +1015,16 @@ def main():
                     _logger.info("Distributing BatchNorm running means and vars")
                 utils.distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
 
+            if (epoch + 1) % args.val_interval != 0:
+                if utils.is_primary(args):
+                    _logger.info("Skipping eval and checkpointing ")
+                if lr_scheduler is not None:
+                    # step LR for next epoch
+                    # careful when using metric dependent lr_scheduler
+                    lr_scheduler.step(epoch + 1, metric=None)
+                # skip validation and metric logic
+                continue
+
             if loader_eval is not None:
                 eval_metrics = validate(
                     model,