Update MobileNetV3 architecture, move some layers from body to head, which only appears when include_top=True.

Pretrained weights are updated and tested.

PiperOrigin-RevId: 401845877

Author: haifeng-jin

keras/applications/applications_test.py

@@ -48,8 +48,8 @@
     (inception_resnet_v2.InceptionResNetV2, 1536),
     (mobilenet.MobileNet, 1024),
     (mobilenet_v2.MobileNetV2, 1280),
-    (mobilenet_v3.MobileNetV3Small, 1024),
-    (mobilenet_v3.MobileNetV3Large, 1280),
+    (mobilenet_v3.MobileNetV3Small, 576),
+    (mobilenet_v3.MobileNetV3Large, 960),
     (densenet.DenseNet121, 1024),
     (densenet.DenseNet169, 1664),
     (densenet.DenseNet201, 1920),
@@ -70,6 +70,23 @@
 
 MODEL_LIST = MODEL_LIST_NO_NASNET + NASNET_LIST
 
+# Parameters for loading weights for MobileNetV3.
+# (class, alpha, minimalistic, include_top)
+MOBILENET_V3_FOR_WEIGHTS = [
+    (mobilenet_v3.MobileNetV3Large, 0.75, False, False),
+    (mobilenet_v3.MobileNetV3Large, 1.0, False, False),
+    (mobilenet_v3.MobileNetV3Large, 1.0, True, False),
+    (mobilenet_v3.MobileNetV3Large, 0.75, False, True),
+    (mobilenet_v3.MobileNetV3Large, 1.0, False, True),
+    (mobilenet_v3.MobileNetV3Large, 1.0, True, True),
+    (mobilenet_v3.MobileNetV3Small, 0.75, False, False),
+    (mobilenet_v3.MobileNetV3Small, 1.0, False, False),
+    (mobilenet_v3.MobileNetV3Small, 1.0, True, False),
+    (mobilenet_v3.MobileNetV3Small, 0.75, False, True),
+    (mobilenet_v3.MobileNetV3Small, 1.0, False, True),
+    (mobilenet_v3.MobileNetV3Small, 1.0, True, True),
+]
+
 
 class ApplicationsTest(tf.test.TestCase, parameterized.TestCase):
 
@@ -93,7 +110,7 @@ def test_application_base(self, app, _):
 
   @parameterized.parameters(*MODEL_LIST)
   def test_application_notop(self, app, last_dim):
-    if 'NASNet' or 'MobileNetV3' in app.__name__:
+    if 'NASNet' in app.__name__:
       only_check_last_dim = True
     else:
       only_check_last_dim = False
@@ -119,10 +136,7 @@ def test_application_variable_input_channels(self, app, last_dim):
       input_shape = (None, None, 1)
     output_shape = _get_output_shape(
         lambda: app(weights=None, include_top=False, input_shape=input_shape))
-    if 'MobileNetV3' in app.__name__:
-      self.assertShapeEqual(output_shape, (None, 1, 1, last_dim))
-    else:
-      self.assertShapeEqual(output_shape, (None, None, None, last_dim))
+    self.assertShapeEqual(output_shape, (None, None, None, last_dim))
     backend.clear_session()
 
     if backend.image_data_format() == 'channels_first':
@@ -131,12 +145,23 @@ def test_application_variable_input_channels(self, app, last_dim):
       input_shape = (None, None, 4)
     output_shape = _get_output_shape(
         lambda: app(weights=None, include_top=False, input_shape=input_shape))
-    if 'MobileNetV3' in app.__name__:
-      self.assertShapeEqual(output_shape, (None, 1, 1, last_dim))
-    else:
-      self.assertShapeEqual(output_shape, (None, None, None, last_dim))
+    self.assertShapeEqual(output_shape, (None, None, None, last_dim))
     backend.clear_session()
 
+  @parameterized.parameters(*MOBILENET_V3_FOR_WEIGHTS)
+  def test_mobilenet_v3_load_weights(
+      self,
+      mobilenet_class,
+      alpha,
+      minimalistic,
+      include_top):
+    mobilenet_class(
+        input_shape=(224, 224, 3),
+        weights='imagenet',
+        alpha=alpha,
+        minimalistic=minimalistic,
+        include_top=include_top)
+
 
 def _get_output_shape(model_fn):
   model = model_fn()

keras/applications/mobilenet_v3.py

@@ -33,17 +33,17 @@
                     'keras-applications/mobilenet_v3/')
 WEIGHTS_HASHES = {
     'large_224_0.75_float': ('765b44a33ad4005b3ac83185abf1d0eb',
-                             'e7b4d1071996dd51a2c2ca2424570e20'),
+                             '40af19a13ebea4e2ee0c676887f69a2e'),
     'large_224_1.0_float': ('59e551e166be033d707958cf9e29a6a7',
-                            '037116398e07f018c0005ffcb0406831'),
+                            '07fb09a5933dd0c8eaafa16978110389'),
     'large_minimalistic_224_1.0_float': ('675e7b876c45c57e9e63e6d90a36599c',
-                                         'a2c33aed672524d1d0b4431808177695'),
+                                         'ec5221f64a2f6d1ef965a614bdae7973'),
     'small_224_0.75_float': ('cb65d4e5be93758266aa0a7f2c6708b7',
-                             '4d2fe46f1c1f38057392514b0df1d673'),
+                             'ebdb5cc8e0b497cd13a7c275d475c819'),
     'small_224_1.0_float': ('8768d4c2e7dee89b9d02b2d03d65d862',
-                            'be7100780f875c06bcab93d76641aa26'),
+                            'd3e8ec802a04aa4fc771ee12a9a9b836'),
     'small_minimalistic_224_1.0_float': ('99cd97fb2fcdad2bf028eb838de69e37',
-                                         '20d4e357df3f7a6361f3a288857b1051'),
+                                         'cde8136e733e811080d9fcd8a252f7e4'),
 }
 
 layers = VersionAwareLayers()
@@ -310,16 +310,16 @@ def MobileNetV3(stack_fn,
       axis=channel_axis, epsilon=1e-3,
       momentum=0.999, name='Conv_1/BatchNorm')(x)
   x = activation(x)
-  x = layers.GlobalAveragePooling2D(keepdims=True)(x)
-  x = layers.Conv2D(
-      last_point_ch,
-      kernel_size=1,
-      padding='same',
-      use_bias=True,
-      name='Conv_2')(x)
-  x = activation(x)
-
   if include_top:
+    x = layers.GlobalAveragePooling2D(keepdims=True)(x)
+    x = layers.Conv2D(
+        last_point_ch,
+        kernel_size=1,
+        padding='same',
+        use_bias=True,
+        name='Conv_2')(x)
+    x = activation(x)
+
     if dropout_rate > 0:
       x = layers.Dropout(dropout_rate)(x)
     x = layers.Conv2D(classes, kernel_size=1, padding='same', name='Logits')(x)
@@ -350,7 +350,7 @@ def MobileNetV3(stack_fn,
       file_name = 'weights_mobilenet_v3_' + model_name + '.h5'
       file_hash = WEIGHTS_HASHES[model_name][0]
     else:
-      file_name = 'weights_mobilenet_v3_' + model_name + '_no_top.h5'
+      file_name = 'weights_mobilenet_v3_' + model_name + '_no_top_v2.h5'
       file_hash = WEIGHTS_HASHES[model_name][1]
     weights_path = data_utils.get_file(
         file_name,