[Fix] VFI: update param_scheduler, hooks, and epoch_base_runner

Author: Yshuo-Li

configs/video_interpolators/cain/cain_b5_g1b32_vimeo90k_triplet.py

@@ -120,11 +120,7 @@
 ]
 test_evaluator = val_evaluator
 
-# 1604 iters == 1 epoch
-epoch_length = 1604
-
-train_cfg = dict(
-    type='IterBasedTrainLoop', max_iters=300_000, val_interval=epoch_length)
+train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=500)
 val_cfg = dict(type='ValLoop')
 test_cfg = dict(type='TestLoop')
 
@@ -138,23 +134,24 @@
 # learning policy
 param_scheduler = dict(
     type='ReduceLR',
-    by_epoch=False,
+    by_epoch=True,
     mode='min',
     factor=0.5,
     patience=5,
-    cooldown=0,
-    verbose=True)
+    cooldown=0)
 
 default_hooks = dict(
     checkpoint=dict(
-        type='CheckpointHook',
-        interval=epoch_length * 4,
-        save_optimizer=True,
-        by_epoch=False),
+        type='CheckpointHook', interval=1, save_optimizer=True, by_epoch=True),
     timer=dict(type='IterTimerHook'),
-    logger=dict(type='LoggerHook', interval=1),
+    logger=dict(type='LoggerHook', interval=100),
     sampler_seed=dict(type='DistSamplerSeedHook'),
     # visualization=dict(type='EditVisualizationHook'),
     param_scheduler=dict(
-        type='ReduceLRSchedulerHook', by_epoch=False, val_metric='MAE'),
+        type='ReduceLRSchedulerHook',
+        by_epoch=True,
+        interval=1,
+        val_metric='MAE'),
 )
+
+log_processor = dict(type='LogProcessor', by_epoch=True)

configs/video_interpolators/flavr/flavr_in4out1_g8b4_vimeo90k_septuplet.py

@@ -131,10 +131,7 @@
 ]
 test_evaluator = val_evaluator
 
-epoch_length = 2020
-
-train_cfg = dict(
-    type='IterBasedTrainLoop', max_iters=1_000_000, val_interval=epoch_length)
+train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=500)
 val_cfg = dict(type='ValLoop')
 test_cfg = dict(type='TestLoop')
 
@@ -146,30 +143,26 @@
     ))
 
 # learning policy
-# 1604 iters == 1 epoch
-total_iters = 1000000
-lr_config = dict(
+param_scheduler = dict(
     type='ReduceLR',
-    by_epoch=False,
+    by_epoch=True,
     mode='min',
     factor=0.5,
     patience=10,
-    cooldown=20,
-    verbose=True)
+    cooldown=20)
 
 default_hooks = dict(
     checkpoint=dict(
-        type='CheckpointHook',
-        interval=epoch_length * 2,
-        save_optimizer=True,
-        by_epoch=False),
+        type='CheckpointHook', interval=1, save_optimizer=True, by_epoch=True),
     timer=dict(type='IterTimerHook'),
     logger=dict(type='LoggerHook', interval=100),
     sampler_seed=dict(type='DistSamplerSeedHook'),
     # visualization=dict(type='EditVisualizationHook'),
     param_scheduler=dict(
         type='ReduceLRSchedulerHook',
-        by_epoch=False,
-        interval=epoch_length,
+        by_epoch=True,
+        interval=1,
         val_metric='MAE'),
 )
+
+log_processor = dict(type='LogProcessor', by_epoch=True)

configs/video_interpolators/tof/base_tof_vfi_nobn_1xb1_vimeo90k_triplet.py

@@ -89,7 +89,7 @@
     ))
 
 # learning policy
-lr_config = dict(
+param_scheduler = dict(
     type='StepLR',
     by_epoch=False,
     gamma=0.5,

mmedit/optimizer/scheduler/reduce_lr_scheduler.py

@@ -47,8 +47,6 @@ class ReduceLR(_ParamScheduler):
         eps (float, optional): Minimal decay applied to lr. If the difference
             between new and old lr is smaller than eps, the update is
             ignored. Default: 1e-8.
-        verbose (bool): If ``True``, prints a message to stdout for
-            each update. Default: ``False``.
         begin (int): Step at which to start updating the learning rate.
             Defaults to 0.
         end (int): Step at which to stop updating the learning rate.
@@ -68,7 +66,6 @@ def __init__(self,
                  cooldown: int = 0,
                  min_lr: float = 0.,
                  eps: float = 1e-8,
-                 verbose: bool = False,
                  **kwargs):
 
         super().__init__(optimizer=optimizer, param_name='lr', **kwargs)
@@ -99,7 +96,6 @@ def __init__(self,
         self.mode_worse = None  # the worse value for the chosen mode
         self.min_lr = min_lr
         self.eps = eps
-        self.verbose = verbose
         self.last_epoch = 0
         self._init_is_better(self.mode)
         self._reset()
@@ -130,11 +126,7 @@ def _get_value(self):
             for group in self.optimizer.param_groups:
                 regular_lr = group[self.param_name]
                 if regular_lr - regular_lr * self.factor > self.eps:
-                    new_lr = max(regular_lr * self.factor, self.min_lr)
-                    if self.verbose:
-                        print(f'Reducing learning rate of {group} from '
-                              f'{regular_lr:.4e} to {new_lr:.4e}.')
-                    regular_lr = new_lr
+                    regular_lr = max(regular_lr * self.factor, self.min_lr)
                 results.append(regular_lr)
             return results
 

mmedit/registry.py

@@ -78,13 +78,15 @@ def register_all_modules(init_default_scope: bool = True) -> None:
             Defaults to True.
     """  # noqa
     import mmedit.datasets  # noqa: F401,F403
+    import mmedit.hooks  # noqa: F401,F403
     import mmedit.metrics  # noqa: F401,F403
     import mmedit.models  # noqa: F401,F403
+    import mmedit.optimizer  # noqa: F401,F403
     import mmedit.transforms  # noqa: F401,F403
 
     if init_default_scope:
         never_created = DefaultScope.get_current_instance() is None \
-                        or not DefaultScope.check_instance_created('mmedit')
+            or not DefaultScope.check_instance_created('mmedit')
         if never_created:
             DefaultScope.get_instance('mmedit', scope_name='mmedit')
             return

tools/dist_train.sh

@@ -16,5 +16,4 @@ python -m torch.distributed.launch \
     --master_port=$PORT \
     $(dirname "$0")/train.py \
     $CONFIG \
-    --seed 0 \
     --launcher pytorch ${@:3}