Merge pull request #968 from dwf/activations_in_conv_seq

Allow elementwise Activations inside ConvolutionalSequence.

blocks/bricks/conv.py

@@ -1,10 +1,9 @@
-from theano import tensor
 from theano.tensor.nnet import conv2d
 from theano.tensor.nnet.abstract_conv import (AbstractConv2d_gradInputs,
                                               get_conv_output_shape)
 from theano.tensor.signal.pool import pool_2d, Pool
 
-from blocks.bricks import Initializable, Feedforward, Sequence
+from blocks.bricks import Initializable, Feedforward, Sequence, Activation
 from blocks.bricks.base import application, Brick, lazy
 from blocks.roles import add_role, FILTER, BIAS
 from blocks.utils import shared_floatx_nans
@@ -496,6 +495,8 @@ class ConvolutionalSequence(Sequence, Initializable, Feedforward):
     layers : list
         List of convolutional bricks (i.e. :class:`Convolutional`,
         :class:`ConvolutionalActivation`, or :class:`Pooling` bricks).
+        :class:`Activation` bricks that operate elementwise can also
+        be included.
     num_channels : int
         Number of input channels in the image. For the first layer this is
         normally 1 for grayscale images and 3 for color (RGB) images. For
@@ -553,13 +554,25 @@ def get_dim(self, name):
         if name == 'input_':
             return ((self.num_channels,) + self.image_size)
         if name == 'output':
-            return self.layers[-1].get_dim(name)
+            last = len(self.layers) - 1
+            while last >= 0:
+                try:
+                    return self.layers[last].get_dim(name)
+                except ValueError:
+                    last -= 1
+            # The output shape of an empty ConvolutionalSequence or one
+            # consisting only of Activations is the input shape.
+            return self.get_dim('input_')
         return super(ConvolutionalSequence, self).get_dim(name)
 
     def _push_allocation_config(self):
         num_channels = self.num_channels
         image_size = self.image_size
         for layer in self.layers:
+            if isinstance(layer, Activation):
+                # Activations operate elementwise; nothing to set.
+                layer.push_allocation_config()
+                continue
             if self.border_mode is not None:
                 layer.border_mode = self.border_mode
             layer.tied_biases = self.tied_biases
@@ -569,7 +582,7 @@ def _push_allocation_config(self):
             layer.use_bias = self.use_bias
 
             # Push input dimensions to children
-            layer._push_allocation_config()
+            layer.push_allocation_config()
 
             # Retrieve output dimensions
             # and set it for next layer

blocks/utils/__init__.py

@@ -81,7 +81,8 @@ def shared_floatx_zeros_matching(shared_variable, name=None, **kwargs):
     name : :obj:`str`, optional
         The name for the shared variable. Defaults to `None`.
     \*\*kwargs
-        Keyword arguments to pass to the :func:`shared_floatx_zeros` function.
+        Keyword arguments to pass to the :func:`shared_floatx_zeros`
+        function.
 
     Returns
     -------

tests/bricks/test_conv.py

@@ -282,6 +282,45 @@ def test_convolutional_sequence():
     assert_allclose(func(x_val), y_val)
 
 
+def test_convolutional_sequence_with_raw_activation():
+    seq = ConvolutionalSequence([Rectifier()], num_channels=4,
+                                image_size=(20, 14))
+    input_ = (((numpy.arange(2 * 4 * 20 * 14)
+                .reshape((2, 4, 20, 14)) % 2) * 2 - 1)
+              .astype(theano.config.floatX))
+    expected_ = input_ * (input_ > 0)
+    x = theano.tensor.tensor4()
+    assert_allclose(seq.apply(x).eval({x: input_}), expected_)
+
+
+def test_convolutional_sequence_with_convolutions_raw_activation():
+    seq = ConvolutionalSequence(
+        [Convolutional(filter_size=(3, 3), num_filters=4),
+         Rectifier(),
+         Convolutional(filter_size=(5, 5), num_filters=3, step=(2, 2)),
+         Tanh()],
+        num_channels=2,
+        image_size=(21, 39))
+    seq.allocate()
+    x = theano.tensor.tensor4()
+    out = seq.apply(x).eval({x: numpy.ones((10, 2, 21, 39),
+                                           dtype=theano.config.floatX)})
+    assert out.shape == (10, 3, 8, 17)
+
+
+def test_convolutional_sequence_activation_get_dim():
+    seq = ConvolutionalSequence([Tanh()], num_channels=9, image_size=(4, 6))
+    seq.allocate()
+    assert seq.get_dim('output') == (9, 4, 6)
+
+    seq = ConvolutionalSequence([Convolutional(filter_size=(7, 7),
+                                               num_filters=5,
+                                               border_mode=(1, 1)),
+                                 Tanh()], num_channels=8, image_size=(8, 11))
+    seq.allocate()
+    assert seq.get_dim('output') == (5, 4, 7)
+
+
 def test_convolutional_activation_use_bias():
     act = ConvolutionalActivation(Rectifier().apply, (3, 3), 5, 4,
                                   image_size=(9, 9), use_bias=False)