Merge pull request #1108 from davidbau/tied_biases

Convolutions should default to tied_biases=True.

blocks/bricks/conv.py

@@ -40,11 +40,11 @@ class Convolutional(LinearLike):
         The border mode to use, see :func:`scipy.signal.convolve2d` for
         details. Defaults to 'valid'.
     tied_biases : bool
-        If ``True``, it indicates that the biases of every filter in this
-        layer should be shared amongst all applications of that filter.
         Setting this to ``False`` will untie the biases, yielding a
         separate bias for every location at which the filter is applied.
-        Defaults to ``False``.
+        If ``True``, it indicates that the biases of every filter in this
+        layer should be shared amongst all applications of that filter.
+        Defaults to ``True``.
 
     """
     # Make it possible to override the implementation of conv2d that gets
@@ -71,7 +71,7 @@ class Convolutional(LinearLike):
     @lazy(allocation=['filter_size', 'num_filters', 'num_channels'])
     def __init__(self, filter_size, num_filters, num_channels, batch_size=None,
                  image_size=(None, None), step=(1, 1), border_mode='valid',
-                 tied_biases=False, **kwargs):
+                 tied_biases=True, **kwargs):
         super(Convolutional, self).__init__(**kwargs)
 
         self.filter_size = filter_size

tests/bricks/test_conv.py

@@ -195,19 +195,58 @@ def test_no_input_size():
                          c.initialize)
 
 
-def test_tied_biases():
+def test_untied_biases():
     x = tensor.tensor4('x')
     num_channels = 4
     num_filters = 3
     batch_size = 5
     filter_size = (3, 3)
     conv = Convolutional(filter_size, num_filters, num_channels,
                          weights_init=Constant(1.), biases_init=Constant(2.),
-                         tied_biases=True)
+                         image_size=(28, 30), tied_biases=False)
+    conv.initialize()
+
+    y = conv.apply(x)
+    func = function([x], y)
+
+    # Untied biases provide a bias for every individual output
+    assert_allclose(conv.b.eval().shape, (3, 26, 28))
+
+    # Untied biases require images of a specific size
+    x_val_1 = numpy.ones((batch_size, num_channels, 28, 30),
+                          dtype=theano.config.floatX)
+
+    assert_allclose(func(x_val_1),
+                    numpy.prod(filter_size) * num_channels *
+                    numpy.ones((batch_size, num_filters, 26, 28)) + 2)
+
+    x_val_2 = numpy.ones((batch_size, num_channels, 23, 19),
+                          dtype=theano.config.floatX)
+
+    def wrongsize():
+        func(x_val_2)
+
+    assert_raises_regexp(ValueError, 'Input dimension mis-match.*',
+                         wrongsize)
+
+
+def test_tied_biases():
+    x = tensor.tensor4('x')
+    num_channels = 4
+    num_filters = 3
+    batch_size = 5
+    filter_size = (3, 3)
+
+    # Tied biases are the default
+    conv = Convolutional(filter_size, num_filters, num_channels,
+                         weights_init=Constant(1.), biases_init=Constant(2.))
     conv.initialize()
     y = conv.apply(x)
     func = function([x], y)
 
+    # Tied biases only provide one bias for each filter
+    assert_allclose(conv.b.eval().shape, (3,))
+
     # Tied biases allows to pass images of different sizes
     x_val_1 = numpy.ones((batch_size, num_channels, 10,
                           12), dtype=theano.config.floatX)