Added support for images whose last dimension is odd. This is more expensive…

theano/sandbox/cuda/fftconv.py

@@ -374,7 +374,7 @@ def mult_and_reduce(input_fft_v, filters_fft_v, input_shape=None,
 
 
 def conv2d_fft(input, filters, image_shape=None, filter_shape=None,
-               border_mode='valid'):
+               border_mode='valid', want_padding_on_last_input_dimension=False):
     """
     expects bc01 input
     performs a valid/full convolution
@@ -383,6 +383,14 @@ def conv2d_fft(input, filters, image_shape=None, filter_shape=None,
     filters: (oc, ic, f0, f1)
 
     border_mode: 'valid' of 'full'
+
+    want_padding_on_last_input_dimension: This code does not support
+    images for which the last dimension (the "width") is odd. To support
+    this, you can either pad your images on your own, or call this function
+    with the `want_padding_on_last_input_dimension` flag set to `True`.
+    This introduces an extra copying step and consumes memory.
+    The return value will still be of the appropriate shape because
+    the padding is trimmed right before the output is returned.
     """
 
     # use symbolic shapes to compute shape info at runtime if not specified
@@ -400,11 +408,19 @@ def conv2d_fft(input, filters, image_shape=None, filter_shape=None,
     # pad filters/image to output shape
     if border_mode == 'valid':
         o0 = i0
-        o1 = i1
+        if want_padding_on_last_input_dimension:
+            o1 = i1 + 1
+            input_padded = T.zeros((b, ic, o0, o1), dtype='float32')
+            input_padded = T.set_subtensor(input_padded[:, :, :i0, :i1],
+                                       input)
+        else:
+            o1 = i1
+            input_padded = input
+
         filters_padded = T.zeros((oc, ic, o0, o1), dtype='float32')
         filters_padded = T.set_subtensor(filters_padded[:, :, :f0, :f1],
                                          filters)
-        input_padded = input
+
     elif border_mode == 'full':
         o0 = i0 + f0 - 1
         o1 = i1 + f1 - 1
@@ -447,7 +463,10 @@ def conv2d_fft(input, filters, image_shape=None, filter_shape=None,
 
     # slice because the convolution was circular, we need it to be valid
     if border_mode == 'valid':
-        output = output_circ[:, :, f0 - 1:, f1 - 1:]
+        if want_padding_on_last_input_dimension:
+            output = output_circ[:, :, f0 - 1:, f1 - 1:(o1-1)]
+        else:
+            output = output_circ[:, :, f0 - 1:, f1 - 1:o1]
     else:
         output = output_circ