Adding support for reduced tail on MobileNetV3.

Author: datumbox

torchvision/models/mobilenetv3.py

@@ -201,19 +201,25 @@ def _mobilenet_v3(
     return model
 
 
-def mobilenet_v3_large(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> MobileNetV3:
+def mobilenet_v3_large(pretrained: bool = False, progress: bool = True, reduced_tail: bool = False,
+                       **kwargs: Any) -> MobileNetV3:
     """
     Constructs a large MobileNetV3 architecture from
     `"Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>`_.
 
     Args:
         pretrained (bool): If True, returns a model pre-trained on ImageNet
         progress (bool): If True, displays a progress bar of the download to stderr
+        reduced_tail (bool): If True, reduces the channel counts of all feature layers
+            between C4 and C5 by 2. It is used to reduce the channel redundancy in the
+            backbone for Detection and Segmentation.
     """
     width_mult = 1.0
     bneck_conf = partial(InvertedResidualConfig, width_mult=width_mult)
     adjust_channels = partial(InvertedResidualConfig.adjust_channels, width_mult=width_mult)
 
+    reduce_divider = 2 if reduced_tail else 1
+
     inverted_residual_setting = [
         bneck_conf(16, 3, 16, 16, False, "RE", 1),
         bneck_conf(16, 3, 64, 24, False, "RE", 2),
@@ -227,28 +233,34 @@ def mobilenet_v3_large(pretrained: bool = False, progress: bool = True, **kwargs
         bneck_conf(80, 3, 184, 80, False, "HS", 1),
         bneck_conf(80, 3, 480, 112, True, "HS", 1),
         bneck_conf(112, 3, 672, 112, True, "HS", 1),
-        bneck_conf(112, 5, 672, 160, True, "HS", 2),
-        bneck_conf(160, 5, 960, 160, True, "HS", 1),
-        bneck_conf(160, 5, 960, 160, True, "HS", 1),
+        bneck_conf(112, 5, 672, 160, True, "HS", 2),  # C4
+        bneck_conf(160 // reduce_divider, 5, 960, 160, True, "HS", 1),
+        bneck_conf(160 // reduce_divider, 5, 960, 160, True, "HS", 1),
     ]
-    last_channel = adjust_channels(1280)
+    last_channel = adjust_channels(1280 // reduce_divider)  # C5
 
     return _mobilenet_v3("mobilenet_v3_large", inverted_residual_setting, last_channel, pretrained, progress, **kwargs)
 
 
-def mobilenet_v3_small(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> MobileNetV3:
+def mobilenet_v3_small(pretrained: bool = False, progress: bool = True, reduced_tail: bool = False,
+                       **kwargs: Any) -> MobileNetV3:
     """
     Constructs a small MobileNetV3 architecture from
     `"Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>`_.
 
     Args:
         pretrained (bool): If True, returns a model pre-trained on ImageNet
         progress (bool): If True, displays a progress bar of the download to stderr
+        reduced_tail (bool): If True, reduces the channel counts of all feature layers
+            between C4 and C5 by 2. It is used to reduce the channel redundancy in the
+            backbone for Detection and Segmentation.
     """
     width_mult = 1.0
     bneck_conf = partial(InvertedResidualConfig, width_mult=width_mult)
     adjust_channels = partial(InvertedResidualConfig.adjust_channels, width_mult=width_mult)
 
+    reduce_divider = 2 if reduced_tail else 1
+
     inverted_residual_setting = [
         bneck_conf(16, 3, 16, 16, True, "RE", 2),
         bneck_conf(16, 3, 72, 24, False, "RE", 2),
@@ -258,10 +270,10 @@ def mobilenet_v3_small(pretrained: bool = False, progress: bool = True, **kwargs
         bneck_conf(40, 5, 240, 40, True, "HS", 1),
         bneck_conf(40, 5, 120, 48, True, "HS", 1),
         bneck_conf(48, 5, 144, 48, True, "HS", 1),
-        bneck_conf(48, 5, 288, 96, True, "HS", 2),
-        bneck_conf(96, 5, 576, 96, True, "HS", 1),
-        bneck_conf(96, 5, 576, 96, True, "HS", 1),
+        bneck_conf(48, 5, 288, 96, True, "HS", 2),  # C4
+        bneck_conf(96 // reduce_divider, 5, 576, 96, True, "HS", 1),
+        bneck_conf(96 // reduce_divider, 5, 576, 96, True, "HS", 1),
     ]
-    last_channel = adjust_channels(1024)
+    last_channel = adjust_channels(1024 // reduce_divider)  # C5
 
     return _mobilenet_v3("mobilenet_v3_small", inverted_residual_setting, last_channel, pretrained, progress, **kwargs)