Speed up median_filter2d

tensorflow_addons/image/filters.py

@@ -174,34 +174,44 @@ def median_filter2d(image,
         if rank == 3:
             image = tf.expand_dims(image, axis=0)
 
+        image_shape = tf.shape(image)
+        batch_size = image_shape[0]
+        height = image_shape[1]
+        width = image_shape[2]
+        channels = image_shape[3]
+
         # Explicitly pad the image
         image = _pad(
             image, filter_shape, mode=padding, constant_values=constant_values)
 
-        floor = (filter_shape[0] * filter_shape[1] + 1) // 2
-        ceil = (filter_shape[0] * filter_shape[1]) // 2 + 1
-
-        def _median_filter2d_single_channel(x):
-            x = tf.expand_dims(x, axis=-1)
-            patches = tf.image.extract_patches(
-                x,
-                sizes=[1, filter_shape[0], filter_shape[1], 1],
-                strides=[1, 1, 1, 1],
-                rates=[1, 1, 1, 1],
-                padding="VALID")
-
-            # Note the returned median is casted back to the original type
-            # Take [5, 6, 7, 8] for example, the median is (6 + 7) / 2 = 3.5
-            # It turns out to be int(6.5) = 6 if the original type is int
-            top = tf.nn.top_k(patches, k=ceil).values
-            median = (top[:, :, :, floor - 1] + top[:, :, :, ceil - 1]) / 2
-            return tf.dtypes.cast(median, x.dtype)
-
-        output = tf.map_fn(
-            _median_filter2d_single_channel,
-            elems=tf.transpose(image, [3, 0, 1, 2]),
-            dtype=image.dtype)
-        output = tf.transpose(output, [1, 2, 3, 0])
+        area = filter_shape[0] * filter_shape[1]
+
+        floor = (area + 1) // 2
+        ceil = area // 2 + 1
+
+        patches = tf.image.extract_patches(
+            image,
+            sizes=[1, filter_shape[0], filter_shape[1], 1],
+            strides=[1, 1, 1, 1],
+            rates=[1, 1, 1, 1],
+            padding="VALID")
+
+        patches = tf.reshape(
+            patches, shape=[batch_size, height, width, area, channels])
+
+        patches = tf.transpose(patches, [0, 1, 2, 4, 3])
+
+        # Note the returned median is casted back to the original type
+        # Take [5, 6, 7, 8] for example, the median is (6 + 7) / 2 = 3.5
+        # It turns out to be int(6.5) = 6 if the original type is int
+        top = tf.nn.top_k(patches, k=ceil).values
+        if area % 2 == 1:
+            median = top[:, :, :, :, floor - 1]
+        else:
+            median = (
+                top[:, :, :, :, floor - 1] + top[:, :, :, :, ceil - 1]) / 2
+
+        output = tf.cast(median, image.dtype)
 
         # Squeeze out the first axis to make sure
         # output has the same dimension with image.