[Imporve] Using train_step instead of forward in PreciseBNHook (#964)

* fix precise BN hook when using MLU

* fix unit tests

mmcls/core/hook/precise_bn_hook.py

@@ -107,7 +107,7 @@ def update_bn_stats(model: nn.Module,
         prog_bar = mmcv.ProgressBar(num_iter)
 
     for data in itertools.islice(loader, num_iter):
-        model(**data)
+        model.train_step(data)
         for i, bn in enumerate(bn_layers):
             running_means[i] += bn.running_mean / num_iter
             running_vars[i] += bn.running_var / num_iter

tests/test_runtime/test_preciseBN_hook.py

@@ -10,6 +10,7 @@ from mmcv.utils.logging import print_log
 from torch.utils.data import DataLoader, Dataset
 
 from mmcls.core.hook import PreciseBNHook
+from mmcls.models.classifiers import BaseClassifier
 
 
 class ExampleDataset(Dataset):
@@ -41,7 +42,7 @@ class BiggerDataset(ExampleDataset):
         return 12
 
 
-class ExampleModel(nn.Module):
+class ExampleModel(BaseClassifier):
 
     def __init__(self):
         super().__init__()
@@ -52,7 +53,17 @@ class ExampleModel(nn.Module):
     def forward(self, imgs, return_loss=False):
         return self.bn(self.conv(imgs))
 
-    def train_step(self, data_batch, optimizer, **kwargs):
+    def simple_test(self, img, img_metas=None, **kwargs):
+        return {}
+
+    def extract_feat(self, img, stage='neck'):
+        return ()
+
+    def forward_train(self, img, gt_label, **kwargs):
+        return {'loss': 0.5}
+
+    def train_step(self, data_batch, optimizer=None, **kwargs):
+        self.forward(**data_batch)
         outputs = {
             'loss': 0.5,
             'log_vars': {
@@ -234,10 +245,8 @@ def test_precise_bn():
     mean = np.mean([np.mean(batch) for batch in imgs_list])
     # bassel correction used in Pytorch, therefore ddof=1
     var = np.mean([np.var(batch, ddof=1) for batch in imgs_list])
-    assert np.equal(mean, np.array(
-        model.bn.running_mean)), (mean, np.array(model.bn.running_mean))
-    assert np.equal(var, np.array(
-        model.bn.running_var)), (var, np.array(model.bn.running_var))
+    assert np.equal(mean, model.bn.running_mean)
+    assert np.equal(var, model.bn.running_var)
 
     @pytest.mark.skipif(
         not torch.cuda.is_available(), reason='requires CUDA support')