time distributed blocks

Author: 0x00b1

keras_resnet/__init__.py

@@ -1,7 +1,6 @@
 from .block import (
     basic,
-    bottleneck,
-    shortcut
+    bottleneck
 )
 
 from .models import (

keras_resnet/block.py renamed to keras_resnet/block/__init__.py

@@ -45,7 +45,7 @@ def f(x):
         y = keras.layers.Conv2D(filters, (3, 3), padding="same", **parameters)(y)
 
         y = keras.layers.BatchNormalization(axis=axis)(y)
-        y = shortcut(x, y)
+        y = _shortcut(x, y)
         y = keras.layers.Activation("relu")(y)
 
         return y
@@ -89,29 +89,29 @@ def f(x):
         y = keras.layers.Conv2D(filters * 4, (1, 1), **parameters)(y)
 
         y = keras.layers.BatchNormalization(axis=axis)(y)
-        y = shortcut(x, y)
+        y = _shortcut(x, y)
         y = keras.layers.Activation("relu")(y)
 
         return y
 
     return f
 
 
-def shortcut(a, b):
+def _shortcut(a, b):
     a_shape = keras.backend.int_shape(a)
     b_shape = keras.backend.int_shape(b)
 
     if keras.backend.image_data_format() == "channels_last":
-        x = int(round(a_shape[1] / b_shape[1]))
-        y = int(round(a_shape[2] / b_shape[2]))
+        x = int(round(a_shape[1] // b_shape[1]))
+        y = int(round(a_shape[2] // b_shape[2]))
 
         if x > 1 or y > 1 or not a_shape[3] == b_shape[3]:
             a = keras.layers.Conv2D(b_shape[3], (1, 1), strides=(x, y), padding="same", **parameters)(a)
 
             a = keras.layers.BatchNormalization(axis=3)(a)
     else:
-        x = int(round(a_shape[2] / b_shape[2]))
-        y = int(round(a_shape[3] / b_shape[3]))
+        x = int(round(a_shape[2] // b_shape[2]))
+        y = int(round(a_shape[3] // b_shape[3]))
 
         if x > 1 or y > 1 or not a_shape[1] == b_shape[1]:
             a = keras.layers.Conv2D(b_shape[1], (1, 1), strides=(x, y), padding="same", **parameters)(a)

keras_resnet/block/temporal.py

@@ -0,0 +1,121 @@
+# -*- coding: utf-8 -*-
+
+"""
+
+keras_resnet.block.temporal
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+This module implements a number of popular time distributed residual blocks.
+
+"""
+
+import keras.layers
+import keras.regularizers
+
+parameters = {
+    "kernel_initializer": "he_normal"
+}
+
+
+def basic(filters, strides=(1, 1), first=False):
+    """
+
+    A time distributed basic block.
+
+    :param filters: the output’s feature space
+    :param strides: the convolution’s stride
+    :param first: whether this is the first instance inside a residual block
+
+    Usage::
+      >>> import keras_resnet.block.temporal
+      >>> keras_resnet.block.temporal.basic(64)
+
+    """
+    def f(x):
+        if keras.backend.image_data_format() == "channels_last":
+            axis = 3
+        else:
+            axis = 1
+
+        y = keras.layers.TimeDistributed(keras.layers.Conv2D(filters, (3, 3), strides=strides, padding="same", **parameters))(x)
+
+        y = keras.layers.TimeDistributed(keras.layers.BatchNormalization(axis=axis))(y)
+        y = keras.layers.TimeDistributed(keras.layers.Activation("relu"))(y)
+
+        y = keras.layers.TimeDistributed(keras.layers.Conv2D(filters, (3, 3), padding="same", **parameters))(y)
+
+        y = keras.layers.TimeDistributed(keras.layers.BatchNormalization(axis=axis))(y)
+        y = _shortcut(x, y)
+        y = keras.layers.TimeDistributed(keras.layers.Activation("relu"))(y)
+
+        return y
+
+    return f
+
+
+def bottleneck(filters, strides=(1, 1), first=False):
+    """
+
+    A time distributed bottleneck block.
+
+    :param filters: the output’s feature space
+    :param strides: the convolution’s stride
+    :param first: whether this is the first instance inside a residual block
+
+    Usage::
+      >>> import keras_resnet.block.temporal
+      >>> keras_resnet.block.temporal.bottleneck(64)
+
+    """
+    def f(x):
+        if keras.backend.image_data_format() == "channels_last":
+            axis = 3
+        else:
+            axis = 1
+
+        if first:
+            y = keras.layers.TimeDistributed(keras.layers.Conv2D(filters, (1, 1), strides=strides, padding="same", **parameters))(x)
+        else:
+            y = keras.layers.TimeDistributed(keras.layers.Conv2D(filters, (3, 3), strides=strides, padding="same", **parameters))(x)
+
+        y = keras.layers.TimeDistributed(keras.layers.BatchNormalization(axis=axis))(y)
+        y = keras.layers.TimeDistributed(keras.layers.Activation("relu"))(y)
+
+        y = keras.layers.TimeDistributed(keras.layers.Conv2D(filters, (3, 3), padding="same", **parameters))(y)
+
+        y = keras.layers.TimeDistributed(keras.layers.BatchNormalization(axis=axis))(y)
+        y = keras.layers.TimeDistributed(keras.layers.Activation("relu"))(y)
+
+        y = keras.layers.TimeDistributed(keras.layers.Conv2D(filters * 4, (1, 1), **parameters))(y)
+
+        y = keras.layers.TimeDistributed(keras.layers.BatchNormalization(axis=axis))(y)
+        y = _shortcut(x, y)
+        y = keras.layers.TimeDistributed(keras.layers.Activation("relu"))(y)
+
+        return y
+
+    return f
+
+
+def _shortcut(a, b):
+    a_shape = keras.backend.int_shape(a)
+    b_shape = keras.backend.int_shape(b)
+
+    if keras.backend.image_data_format() == "channels_last":
+        x = int(round(a_shape[1] // b_shape[1]))
+        y = int(round(a_shape[2] // b_shape[2]))
+
+        if x > 1 or y > 1 or not a_shape[3] == b_shape[3]:
+            a = keras.layers.TimeDistributed(keras.layers.Conv2D(b_shape[3], (1, 1), strides=(x, y), padding="same", **parameters))(a)
+
+            a = keras.layers.TimeDistributed(keras.layers.BatchNormalization(axis=3))(a)
+    else:
+        x = int(round(a_shape[2] // b_shape[2]))
+        y = int(round(a_shape[3] // b_shape[3]))
+
+        if x > 1 or y > 1 or not a_shape[1] == b_shape[1]:
+            a = keras.layers.TimeDistributed(keras.layers.Conv2D(b_shape[1], (1, 1), strides=(x, y), padding="same", **parameters))(a)
+
+            a = keras.layers.TimeDistributed(keras.layers.BatchNormalization(axis=1))(a)
+
+    return keras.layers.add([a, b])

keras_resnet/models.py

@@ -42,10 +42,12 @@ def __init__(self, inputs, blocks, block):
         else:
             axis = 1
 
-        x = keras.layers.Conv2D(64, (7, 7), strides=(2, 2), padding="same")(inputs)
+        x = keras.layers.Conv2D(64, (7, 7), strides=(2, 2), padding="same")(
+            inputs)
         x = keras.layers.BatchNormalization(axis=axis)(x)
         x = keras.layers.Activation("relu")(x)
-        x = keras.layers.MaxPooling2D((3, 3), strides=(2, 2), padding="same")(x)
+        x = keras.layers.MaxPooling2D((3, 3), strides=(2, 2), padding="same")(
+            x)
 
         features = 64
 
@@ -94,6 +96,7 @@ class ResNet18(ResNet):
         >>> model.compile("adam", "categorical_crossentropy", ["accuracy"])
 
     """
+
     def __init__(self, inputs):
         block = keras_resnet.block.basic
 
@@ -115,6 +118,7 @@ class ResNet34(ResNet):
         >>> model.compile("adam", "categorical_crossentropy", ["accuracy"])
 
     """
+
     def __init__(self, inputs):
         block = keras_resnet.block.basic
 
@@ -136,6 +140,7 @@ class ResNet50(ResNet):
         >>> model.compile("adam", "categorical_crossentropy", ["accuracy"])
 
     """
+
     def __init__(self, inputs):
         block = keras_resnet.block.bottleneck
 
@@ -157,6 +162,7 @@ class ResNet101(ResNet):
         >>> model.compile("adam", "categorical_crossentropy", ["accuracy"])
 
     """
+
     def __init__(self, inputs):
         block = keras_resnet.block.bottleneck
 
@@ -178,6 +184,7 @@ class ResNet152(ResNet):
         >>> model.compile("adam", "categorical_crossentropy", ["accuracy"])
 
     """
+
     def __init__(self, inputs):
         block = keras_resnet.block.bottleneck
 
@@ -199,6 +206,7 @@ class ResNet200(ResNet):
         >>> model.compile("adam", "categorical_crossentropy", ["accuracy"])
 
     """
+
     def __init__(self, inputs):
         block = keras_resnet.block.bottleneck