Remove experiemental conv_2d_offset code.

theano/tensor/nnet/conv.py

@@ -83,93 +83,6 @@ def conv2d(input, filters, image_shape=None, filter_shape=None,
 
     return op(input, filters)
 
-def conv2d_offset(input, filters, image_shape=None, filter_shape=None,
-                border_mode='valid', subsample=(1,1), offsets = [], **kargs):
-    """
-    Build a graph for nnet convolutions with subsampling and offsetting.
-
-    Refer to conv2d for a general explanation of the context.
-
-    *Subsampling* means, we only compute convolutions at a fraction of the 
-    sites. *Offsetting* convolutions in the context of subsampling means not
-    computing all of them at the same sites (starting in upper left corner).
-    This function allows offsets to be used.
-
-    Most parameters are shared with conv2d, except:
-    :param offsets: list of 2-tuples that specify sites wrt the subsampling
-    scheme. The filters are split evenly accross the different sites. For 
-    example with subsample (2,2) we can use the sites [(0,0), (0,1), (1,0),
-    (1,1)]. An empty list is interpreted as all offsets.
-    """
-    # There should be subsampling for offsets to make any sense.
-    if not (subsample[0] > 1 or subsample[1] > 1):
-        raise ValueError('conv2d_offset requires subsampling.')
-
-    # Haven't thought about this case.
-    if numpy.any(numpy.array(subsample) > filter_shape[2:]):
-        raise ValueError('conv2d_offset subsample greater than filter shape. Not supported?')
-
-    # No offsets specified is interpreted as all offsets.
-    if len(offsets) == 0:
-        offsets = []
-        for i in range(subsample[0]):
-            for j in range(subsample[1]):
-              offsets.append((i,j))
-
-    # Find the largest offsets in both image dimensions. Used to determine the
-    # size of the image used.
-    max_offset = list(offsets[0])
-
-    for offset in offsets:
-        if offset[0] > max_offset[0]:
-            max_offset[0] = offset[0]
-        if offset[1] > max_offset[1]:
-            max_offset[1] = offset[1]
-
-    if not (max_offset[0] < subsample[0] and max_offset[1] < subsample[1]):
-        raise ValueError('conv2d_offset: invalid offset sites.')
-
-    # Determine the reduced size of input so all feature maps get an input of 
-    # the same size.
-    sub_image_shape = list(image_shape)
-    sub_image_shape[2] -= max_offset[0]
-    sub_image_shape[3] -= max_offset[1]
-
-    # Determine number of filters per offset position.
-    if (filter_shape[0] % len(offsets)) != 0:
-        print 'nfilts ', filter_shape[0], ' noffsets ', len(offsets)
-        raise ValueError('conv2d_offset: invalid number of filters wrt offsets.')
-    n_filters = filter_shape[0] / len(offsets) 
-    sub_filter_shape = list(filter_shape)
-    sub_filter_shape[0] = n_filters
-
-    # Call conv2d at each offset using same fraction of kernels.
-    outputs = []
-    for i, offset in enumerate(offsets):
-
-        # Crop the input so all offsets get an input of the same size.
-        sub_input = input[:, :, offset[0]:sub_image_shape[2] + offset[0], 
-                      offset[1]:sub_image_shape[3] + offset[1]]
-
-        # Grab part of the filters.
-        sub_filters = filters[i*n_filters:(i+1)*n_filters]
-
-        out = conv2d(sub_input, sub_filters, sub_image_shape, sub_filter_shape,
-                    border_mode=border_mode, subsample=subsample, **kargs)
-
-        outputs.append(out)
-
-    # Join the outputs on the leading axis.
-    if len(outputs) > 1:
-        output = tensor.join(1, *outputs)
-    else:
-        output = outputs[0] 
-
-    outshp = ConvOp.getOutputShape(sub_image_shape[2:], filter_shape[2:], subsample, border_mode)
-
-    return [output, outshp]
-
-
 
 class ConvOp(Op):
     """