Added Subpixel convolution layer test

keras_contrib/backend/tensorflow_backend.py

@@ -101,7 +101,7 @@ def extract_image_patches(X, ksizes, ssizes, border_mode="same", dim_ordering="t
     return patches
 
 
-def depth_to_space(input, scale, **kwargs):
+def depth_to_space(input, scale):
     ''' Uses phase shift algorithm to convert channels/depth for spatial resolution '''
 
     return tf.depth_to_space(input, scale)

keras_contrib/layers/convolutional.py

@@ -231,10 +231,14 @@ get_custom_objects().update({"Deconv3D": Deconv3D})
 
 class SubPixelUpscaling(Layer):
 
-    def __init__(self, scale_factor=2, **kwargs):
+    def __init__(self, scale_factor=2, dim_ordering='default', **kwargs):
         super(SubPixelUpscaling, self).__init__(**kwargs)
 
         self.scale_factor = scale_factor
+        self.dim_ordering = dim_ordering
+
+        if self.dim_ordering == 'default':
+            self.dim_ordering = K.image_dim_ordering()
 
     def build(self, input_shape):
         pass
@@ -244,7 +248,7 @@ class SubPixelUpscaling(Layer):
         return y
 
     def get_output_shape_for(self, input_shape):
-        if K.image_dim_ordering() == "th":
+        if self.dim_ordering == 'th':
             b, k, r, c = input_shape
             return (b, k // (self.scale_factor ** 2), r * self.scale_factor, c * self.scale_factor)
         else:
@@ -252,6 +256,9 @@ class SubPixelUpscaling(Layer):
             return (b, r * self.scale_factor, c * self.scale_factor, k // (self.scale_factor ** 2))
 
     def get_config(self):
-        config = {'scale_factor': self.scale_factor}
+        config = {'scale_factor': self.scale_factor,
+                  'dim_ordering': self.dim_ordering}
         base_config = super(SubPixelUpscaling, self).get_config()
         return dict(list(base_config.items()) + list(config.items()))
+
+get_custom_objects().update({'SubPixelUpscaling': SubPixelUpscaling})

tests/keras_contrib/layers/test_convolutional.py

@@ -1,5 +1,6 @@
 import pytest
 import numpy as np
+import itertools
 from numpy.testing import assert_allclose
 
 from keras.utils.test_utils import layer_test, keras_test
@@ -76,5 +77,28 @@ def test_deconvolution_3d():
                            input_shape=(nb_samples, stack_size, kernel_dim1, kernel_dim2, kernel_dim3))
 
 
+@keras_test
+def test_sub_pixel_upscaling():
+    nb_samples = 2
+    nb_row = 16
+    nb_col = 16
+
+    for scale_factor in [2, 3, 4]:
+        input_data = np.random.random((nb_samples, 4 * (scale_factor ** 2), nb_row, nb_col))
+
+        if K.image_dim_ordering() == 'tf':
+            input_data = input_data.transpose((0, 2, 3, 1))
+
+        input_tensor = K.variable(input_data)
+        expected_output = K.eval(KC.depth_to_space(input_tensor, scale=scale_factor))
+
+        layer_test(convolutional.SubPixelUpscaling,
+                   kwargs={'scale_factor': scale_factor},
+                   input_data=input_data,
+                   expected_output=expected_output,
+                   expected_output_dtype=K.floatx(),
+                   fixed_batch_size=False)
+
+
 if __name__ == '__main__':
     pytest.main([__file__])