fix dynamic_shape bug of SyncRandomSizeHook (open-mmlab#6144)

* fix SyncRandomSizeHook dynamic_shape bug

* fix unittest

* fix unittest

* update comments

* fix scale error

* update docstr

configs/yolox/yolox_s_8x8_300e_coco.py

@@ -129,7 +129,6 @@
         type='SyncRandomSizeHook',
         ratio_range=(14, 26),
         img_scale=img_scale,
-        interval=interval,
         priority=48),
     dict(
         type='SyncNormHook',

configs/yolox/yolox_tiny_8x8_300e_coco.py

@@ -66,7 +66,6 @@
         type='SyncRandomSizeHook',
         ratio_range=(10, 20),
         img_scale=img_scale,
-        interval=interval,
         priority=48),
     dict(
         type='SyncNormHook',

mmdet/core/hook/sync_random_size_hook.py

@@ -1,5 +1,6 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 import random
+import warnings
 
 import torch
 from mmcv.runner import get_dist_info
@@ -9,34 +10,48 @@
 
 @HOOKS.register_module()
 class SyncRandomSizeHook(Hook):
-    """Change and synchronize the random image size across ranks, currently
-    used in YOLOX.
+    """Change and synchronize the random image size across ranks.
+    SyncRandomSizeHook is deprecated, please use Resize pipeline
+    to achieve similar functions. Such as `dict(type='Resize', img_scale=[(448,
+    448), (832, 832)], multiscale_mode='range', keep_ratio=True)`.
+
+    Note: Due to the multi-process dataloader, its behavior is different
+    from YOLOX's official implementation, the official is to change the
+    size every fixed iteration interval and what we achieved is a fixed
+    epoch interval.
 
     Args:
         ratio_range (tuple[int]): Random ratio range. It will be multiplied
             by 32, and then change the dataset output image size.
             Default: (14, 26).
         img_scale (tuple[int]): Size of input image. Default: (640, 640).
-        interval (int): The interval of change image size. Default: 10.
+        interval (int): The epoch interval of change image size. Default: 1.
         device (torch.device | str): device for returned tensors.
             Default: 'cuda'.
     """
 
     def __init__(self,
                  ratio_range=(14, 26),
                  img_scale=(640, 640),
-                 interval=10,
+                 interval=1,
                  device='cuda'):
+        warnings.warn('DeprecationWarning: SyncRandomSizeHook is deprecated. '
+                      'Please use Resize pipeline to achieve similar '
+                      'functions. Due to the multi-process dataloader, '
+                      'its behavior is different from YOLOX\'s official '
+                      'implementation, the official is to change the size '
+                      'every fixed iteration interval and what we achieved '
+                      'is a fixed epoch interval.')
         self.rank, world_size = get_dist_info()
         self.is_distributed = world_size > 1
         self.ratio_range = ratio_range
         self.img_scale = img_scale
         self.interval = interval
         self.device = device
 
-    def after_train_iter(self, runner):
+    def after_train_epoch(self, runner):
         """Change the dataset output image size."""
-        if self.ratio_range is not None and (runner.iter +
+        if self.ratio_range is not None and (runner.epoch +
                                              1) % self.interval == 0:
             # Due to DDP and DP get the device behavior inconsistent,
             # so we did not get the device from runner.model.

mmdet/datasets/dataset_wrappers.py

@@ -368,8 +368,6 @@ def __getitem__(self, idx):
 
             if 'mix_results' in results:
                 results.pop('mix_results')
-            if 'img_scale' in results:
-                results.pop('img_scale')
 
         return results
 

tests/test_utils/test_hook.py

@@ -12,7 +12,7 @@
 from mmcv.runner import (CheckpointHook, IterTimerHook, PaviLoggerHook,
                          build_runner)
 from torch.nn.init import constant_
-from torch.utils.data import DataLoader
+from torch.utils.data import DataLoader, Dataset
 
 from mmdet.core.hook import ExpMomentumEMAHook, YOLOXLrUpdaterHook
 from mmdet.core.hook.sync_norm_hook import SyncNormHook
@@ -254,12 +254,31 @@ def test_sync_random_size_hook():
     # Only used to prevent program errors
     SyncRandomSizeHook()
 
-    loader = DataLoader(torch.ones((5, 2)))
+    class DemoDataset(Dataset):
+
+        def __getitem__(self, item):
+            return torch.ones(2)
+
+        def __len__(self):
+            return 5
+
+        def update_dynamic_scale(self, dynamic_scale):
+            pass
+
+    loader = DataLoader(DemoDataset())
     runner = _build_demo_runner()
-    runner.register_hook_from_cfg(dict(type='SyncRandomSizeHook'))
+    runner.register_hook_from_cfg(
+        dict(type='SyncRandomSizeHook', device='cpu'))
     runner.run([loader, loader], [('train', 1), ('val', 1)])
     shutil.rmtree(runner.work_dir)
 
+    if torch.cuda.is_available():
+        runner = _build_demo_runner()
+        runner.register_hook_from_cfg(
+            dict(type='SyncRandomSizeHook', device='cuda'))
+        runner.run([loader, loader], [('train', 1), ('val', 1)])
+        shutil.rmtree(runner.work_dir)
+
 
 @pytest.mark.parametrize('set_loss', [
     dict(set_loss_nan=False, set_loss_inf=False),