FPN

keras_resnet/models/_feature_pyramid_2d.py

@@ -18,14 +18,14 @@
 
 class FPN2D(keras.Model):
     def __init__(
-        self,
-        inputs,
-        blocks,
-        block,
-        freeze_bn=True,
-        numerical_names=None,
-        *args,
-        **kwargs
+            self,
+            inputs,
+            blocks,
+            block,
+            freeze_bn=True,
+            numerical_names=None,
+            *args,
+            **kwargs
     ):
         if keras.backend.image_data_format() == "channels_last":
             axis = 3
@@ -36,7 +36,6 @@ def __init__(
             numerical_names = [True] * len(blocks)
 
         x = keras.layers.Conv2D(64, (7, 7), strides=(2, 2), use_bias=False, name="conv1", padding="same")(inputs)
-        # x = keras.layers.Lambda(lambda x_in: tensorflow.contrib.layers.group_norm(x_in, groups=1, channels_axis=axis))(x)
         x = keras_resnet.layers.BatchNormalization(axis=axis, epsilon=1e-5, freeze=freeze_bn, name="bn_conv1")(x)
         x = keras.layers.Activation("relu", name="conv1_relu")(x)
         x = keras.layers.MaxPooling2D((3, 3), strides=(2, 2), padding="same", name="pool1")(x)
@@ -72,18 +71,9 @@ def __init__(
         upsampled_p5 = keras.layers.UpSampling2D(
             interpolation="bilinear",
             name="p5_upsampled",
-            # size=(c4.shape[1], c4.shape[2]) # NOT THE SHAPE ! THE STRIDE
             size=(2, 2)
         )(pyramid_5)
 
-        pyramid_5 = keras.layers.Conv2D(
-            filters=256,
-            kernel_size=3,
-            strides=1,
-            padding="same",
-            name="p5"
-        )(pyramid_5)
-
         pyramid_4 = keras.layers.Conv2D(
             filters=256,
             kernel_size=1,
@@ -99,7 +89,6 @@ def __init__(
         upsampled_p4 = keras.layers.UpSampling2D(
             interpolation="bilinear",
             name="p4_upsampled",
-            # size=(c3.shape[1], c3.shape[2])
             size=(2, 2)
         )(pyramid_4)
 
@@ -126,7 +115,6 @@ def __init__(
         upsampled_p3 = keras.layers.UpSampling2D(
             interpolation="bilinear",
             name="p3_upsampled",
-            # size=(c3.shape[1], c3.shape[2])
             size=(2, 2)
         )(pyramid_3)
 
@@ -158,13 +146,7 @@ def __init__(
             name="p2"
         )(pyramid_2)
 
-        pyramid_6 = keras.layers.Conv2D(
-            filters=256,
-            kernel_size=3,
-            strides=2,
-            padding="same",
-            name="p6"
-        )(c5)
+        pyramid_6 = keras.layers.MaxPooling2D(strides=2, name="p6")(pyramid_5)
 
         outputs = [
             pyramid_2,
@@ -182,26 +164,29 @@ def __init__(
         )
 
 
-class FPN2D18(FPN2D):
+class FPN2D50(FPN2D):
     def __init__(self, inputs, blocks=None, *args, **kwargs):
         if blocks is None:
-            blocks = [2, 2, 2, 2]
+            blocks = [3, 4, 6, 3]
 
-        super(FPN2D18, self).__init__(
+        numerical_names = [False, False, False, False]
+
+        super(FPN2D50, self).__init__(
             inputs,
             blocks,
-            block=keras_resnet.blocks.basic_2d,
+            numerical_names=numerical_names,
+            block=keras_resnet.blocks.bottleneck_2d,
             *args,
             **kwargs
         )
 
 
-class FPN2D34(FPN2D):
+class FPN2D18(FPN2D):
     def __init__(self, inputs, blocks=None, *args, **kwargs):
         if blocks is None:
-            blocks = [3, 4, 6, 3]
+            blocks = [2, 2, 2, 2]
 
-        super(FPN2D34, self).__init__(
+        super(FPN2D18, self).__init__(
             inputs,
             blocks,
             block=keras_resnet.blocks.basic_2d,
@@ -210,18 +195,15 @@ def __init__(self, inputs, blocks=None, *args, **kwargs):
         )
 
 
-class FPN2D50(FPN2D):
+class FPN2D34(FPN2D):
     def __init__(self, inputs, blocks=None, *args, **kwargs):
         if blocks is None:
             blocks = [3, 4, 6, 3]
 
-        numerical_names = [False, False, False, False]
-
-        super(FPN2D50, self).__init__(
+        super(FPN2D34, self).__init__(
             inputs,
             blocks,
-            numerical_names=numerical_names,
-            block=keras_resnet.blocks.bottleneck_2d,
+            block=keras_resnet.blocks.basic_2d,
             *args,
             **kwargs
         )