[Enhance] Enable to toggle whether Gem Pooling is trainable or not. (#1246)

* Enable to toggle whether Gem Pooling is trainable or not.

* Add test case whether Gem Pooling is trainable or not.

* Enable to toggle whether Gem Pooling is trainable or not by requires_grad

---------

Co-authored-by: Yusuke Fujimoto <yusuke.fujimoto@rist.co.jp>

mmcls/models/necks/gem.py

@@ -22,20 +22,20 @@ class GeneralizedMeanPooling(nn.Module):
     has a batch dimension of size 1, which can lead to unexpected errors.
 
     Args:
-        p (float): Parameter value.
+        p (float): Parameter value. Defaults to 3.
-            Default: 3.
+        eps (float): epsilon. Defaults to 1e-6.
-        eps (float): epsilon.
+        clamp (bool): Use clamp before pooling. Defaults to True
-            Default: 1e-6
+        p_trainable (bool): Toggle whether Parameter p is trainable or not.
-        clamp (bool): Use clamp before pooling.
+            Defaults to True.
-            Default: True
     """
 
-    def __init__(self, p=3., eps=1e-6, clamp=True):
+    def __init__(self, p=3., eps=1e-6, clamp=True, p_trainable=True):
         assert p >= 1, "'p' must be a value greater than 1"
         super(GeneralizedMeanPooling, self).__init__()
-        self.p = Parameter(torch.ones(1) * p)
+        self.p = Parameter(torch.ones(1) * p, requires_grad=p_trainable)
         self.eps = eps
         self.clamp = clamp
+        self.p_trainable = p_trainable
 
     def forward(self, inputs):
         if isinstance(inputs, tuple):

tests/test_models/test_necks.py

@@ -44,6 +44,10 @@ def test_gem_neck():
 
     # test gem_neck
     neck = GeneralizedMeanPooling()
+
+    # default p is trainable
+    assert neck.p.requires_grad
+
     # batch_size, num_features, feature_size(2)
     fake_input = torch.rand(1, 16, 24, 24)
 
@@ -61,6 +65,19 @@ def test_gem_neck():
     assert output[0].shape == (1, 8)
     assert output[1].shape == (1, 16)
 
+    # test gem_neck with p_trainable=False
+    neck = GeneralizedMeanPooling(p_trainable=False)
+
+    # p is not trainable
+    assert not neck.p.requires_grad
+
+    # batch_size, num_features, feature_size(2)
+    fake_input = torch.rand(1, 16, 24, 24)
+
+    output = neck(fake_input)
+    # batch_size, num_features
+    assert output.shape == (1, 16)
+
     with pytest.raises(AssertionError):
         # p must be a value greater then 1
         GeneralizedMeanPooling(p=0.5)