Add dilation support

bedanar · Oct 7, 2016 · c595e98 · c595e98
1 parent 06d90ec
commit c595e98
Show file tree

Hide file tree

Showing 91 changed files with 64 additions and 43 deletions.
diff --git a/README.md b/README.md
@@ -53,6 +53,18 @@ A technical report on convolution arithmetic in the context of deep learning.
     <td>Padding, strides, transposed (odd)</td>
     <td></td>
   </tr>
+  <tr>
+    <td><img src="gif/dilation.gif"></td>
+    <td></td>
+    <td></td>
+    <td></td>
+  </tr>
+  <tr>
+    <td>No padding, no stride, dilation</td>
+    <td></td>
+    <td></td>
+    <td></td>
+  </tr>
 </table>
 
 ## Generating the Makefile

diff --git a/bin/generate_makefile b/bin/generate_makefile
@@ -8,45 +8,47 @@ numerical_files = 'bin/produce_figure templates/numerical_figure.txt'
 
 animations = (
     ('no_padding_no_strides',
-     ('arithmetic', 4, 2, 0, 3, 1, 'convolution', False)),
+     ('arithmetic', 4, 2, 0, 3, 1, 1, 'convolution', False)),
     ('no_padding_no_strides_transposed',
-     ('arithmetic', 4, 2, 0, 3, 1, 'convolution', True)),
+     ('arithmetic', 4, 2, 0, 3, 1, 1, 'convolution', True)),
     ('arbitrary_padding_no_strides',
-     ('arithmetic', 5, 6, 2, 4, 1, 'convolution', False)),
+     ('arithmetic', 5, 6, 2, 4, 1, 1, 'convolution', False)),
     ('arbitrary_padding_no_strides_transposed',
-     ('arithmetic', 5, 6, 2, 4, 1, 'convolution', True)),
+     ('arithmetic', 5, 6, 2, 4, 1, 1, 'convolution', True)),
     ('same_padding_no_strides',
-     ('arithmetic', 5, 5, 1, 3, 1, 'convolution', False)),
+     ('arithmetic', 5, 5, 1, 3, 1, 1, 'convolution', False)),
     ('same_padding_no_strides_transposed',
-     ('arithmetic', 5, 5, 1, 3, 1, 'convolution', True)),
+     ('arithmetic', 5, 5, 1, 3, 1, 1, 'convolution', True)),
     ('full_padding_no_strides',
-     ('arithmetic', 5, 7, 2, 3, 1, 'convolution', False)),
+     ('arithmetic', 5, 7, 2, 3, 1, 1, 'convolution', False)),
     ('full_padding_no_strides_transposed',
-     ('arithmetic', 5, 7, 2, 3, 1, 'convolution', True)),
+     ('arithmetic', 5, 7, 2, 3, 1, 1, 'convolution', True)),
     ('no_padding_strides',
-     ('arithmetic', 5, 2, 0, 3, 2, 'convolution', False)),
+     ('arithmetic', 5, 2, 0, 3, 2, 1, 'convolution', False)),
     ('no_padding_strides_transposed',
-     ('arithmetic', 5, 2, 0, 3, 2, 'convolution', True)),
+     ('arithmetic', 5, 2, 0, 3, 2, 1, 'convolution', True)),
     ('padding_strides',
-     ('arithmetic', 5, 3, 1, 3, 2, 'convolution', False)),
+     ('arithmetic', 5, 3, 1, 3, 2, 1, 'convolution', False)),
     ('padding_strides_transposed',
-     ('arithmetic', 5, 3, 1, 3, 2, 'convolution', True)),
+     ('arithmetic', 5, 3, 1, 3, 2, 1, 'convolution', True)),
     ('padding_strides_odd',
-     ('arithmetic', 6, 3, 1, 3, 2, 'convolution', False)),
+     ('arithmetic', 6, 3, 1, 3, 2, 1, 'convolution', False)),
     ('padding_strides_odd_transposed',
-     ('arithmetic', 6, 3, 1, 3, 2, 'convolution', True)),
+     ('arithmetic', 6, 3, 1, 3, 2, 1, 'convolution', True)),
+    ('dilation',
+     ('arithmetic', 7, 3, 0, 3, 1, 2, 'convolution', False)),
     ('numerical_no_padding_no_strides',
-     ('numerical', 5, 3, 0, 3, 1, 'convolution', False)),
+     ('numerical', 5, 3, 0, 3, 1, 1, 'convolution', False)),
     ('numerical_padding_strides',
-     ('numerical', 5, 3, 1, 3, 2, 'convolution', False)),
+     ('numerical', 5, 3, 1, 3, 2, 1, 'convolution', False)),
     ('numerical_average_pooling',
-     ('numerical', 5, 3, 0, 3, 1, 'average', False)),
+     ('numerical', 5, 3, 0, 3, 1, 1, 'average', False)),
     ('numerical_max_pooling',
-     ('numerical', 5, 3, 0, 3, 1, 'max', False)),
+     ('numerical', 5, 3, 0, 3, 1, 1, 'max', False)),
 )
 
 fields = ('type', 'input-size', 'output-size', 'padding', 'kernel-size',
-          'stride', 'mode', 'transposed')
+          'stride', 'dilation', 'mode', 'transposed')
 animations = dict([(name, dict(zip(fields, config)))
                    for name, config in animations])
 
@@ -115,6 +117,8 @@ def make_pdf_section():
                 '--padding={} '.format(config['padding']) +
                 '--kernel-size={} '.format(config['kernel-size']) +
                 '--stride={} '.format(config['stride']) +
+                ('--dilation={} '.format(config['dilation'])
+                 if config['dilation'] != 1 else '' ) +
                 ('--mode={} '.format(config['mode'])
                  if config['type'] == 'numerical' else '') +
                 ('--transposed\n' if config['transposed'] else '\n'))

diff --git a/bin/produce_figure b/bin/produce_figure
@@ -12,7 +12,7 @@ numpy.random.seed(1234)
 
 
 def make_numerical_tex_string(step, input_size, output_size, padding,
-                              kernel_size, stride, mode):
+                              kernel_size, stride, dilation, mode):
     """Creates a LaTeX string for a numerical convolution animation.
 
     Parameters
@@ -42,6 +42,8 @@ def make_numerical_tex_string(step, input_size, output_size, padding,
     if mode not in ('convolution', 'average', 'max'):
         raise ValueError("wrong convolution mode, choices are 'convolution', "
                          "'average' or 'max'")
+    if dilation != 1:
+        raise ValueError("Only a dilation of 1 is currently supported for numerical output")
     max_steps = output_size ** 2
     if step >= max_steps:
         raise ValueError('step {} out of bounds (there are '.format(step) +
@@ -116,7 +118,7 @@ def make_numerical_tex_string(step, input_size, output_size, padding,
 
 
 def make_arithmetic_tex_string(step, input_size, output_size, padding,
-                               kernel_size, stride, transposed):
+                               kernel_size, stride, dilation, transposed):
     """Creates a LaTeX string for a convolution arithmetic animation.
 
     Parameters
@@ -133,6 +135,8 @@ def make_arithmetic_tex_string(step, input_size, output_size, padding,
         Convolution kernel size.
     stride : int
         Convolution stride.
+    dilation: int
+        Input Dilation
     transposed : bool
         If ``True``, generate strings for the transposed convolution
         animation.
@@ -144,6 +148,7 @@ def make_arithmetic_tex_string(step, input_size, output_size, padding,
         animation.
 
     """
+    kernel_size = (kernel_size - 1)*dilation + 1
     if transposed:
         input_size, output_size, padding, spacing, stride = (
             output_size, input_size, kernel_size - 1 - padding, stride, 1)
@@ -177,22 +182,15 @@ def make_arithmetic_tex_string(step, input_size, output_size, padding,
                                  padding + spacing * j + 1)
             for i, j in itertools.product(range(input_size),
                                           range(input_size))),
-        'INPUT_GRID_FROM': '{},{}'.format(
-            stride * offset_x, y_adjustment + stride * offset_y),
-        'INPUT_GRID_TO': '{},{}'.format(
-            stride * offset_x + kernel_size,
-            y_adjustment + stride * offset_y + kernel_size),
-        'INPUT_BOTTOM_LEFT': '{},{}'.format(
-            stride * offset_x, y_adjustment + stride * offset_y),
-        'INPUT_BOTTOM_RIGHT': '{},{}'.format(
-            stride * offset_x + kernel_size,
+        'INPUT_GRID_FROM_X': '{}'.format(
+            stride * offset_x),
+        'INPUT_GRID_FROM_Y': '{}'.format(
             y_adjustment + stride * offset_y),
-        'INPUT_TOP_LEFT': '{},{}'.format(
-            stride * offset_x,
-            y_adjustment + stride * offset_y + kernel_size),
-        'INPUT_TOP_RIGHT': '{},{}'.format(
-            stride * offset_x + kernel_size,
+        'INPUT_GRID_TO_X': '{}'.format(
+            stride * offset_x + kernel_size),
+        'INPUT_GRID_TO_Y': '{}'.format(
             y_adjustment + stride * offset_y + kernel_size),
+        'DILATION': '{}'.format(dilation),
         'OUTPUT_BOTTOM_LEFT': '{},{}'.format(offset_x, offset_y),
         'OUTPUT_BOTTOM_RIGHT': '{},{}'.format(offset_x + 1, offset_y),
         'OUTPUT_TOP_LEFT': '{},{}'.format(offset_x, offset_y + 1),
@@ -239,6 +237,8 @@ if __name__ == "__main__":
                                help="kernel size")
     parent_parser.add_argument("-s", "--stride", type=int, default=1,
                                help="stride")
+    parent_parser.add_argument("-d", "--dilation", type=int, default=1,
+                               help="dilation")
 
     subparser = subparsers.add_parser('arithmetic', parents=[parent_parser],
                                       help='convolution arithmetic animation')

diff --git a/gif/arbitrary_padding_no_strides.gif b/gif/arbitrary_padding_no_strides.gif
diff --git a/gif/arbitrary_padding_no_strides_transposed.gif b/gif/arbitrary_padding_no_strides_transposed.gif
diff --git a/gif/full_padding_no_strides_transposed.gif b/gif/full_padding_no_strides_transposed.gif
diff --git a/gif/no_padding_no_strides.gif b/gif/no_padding_no_strides.gif
diff --git a/gif/no_padding_no_strides_transposed.gif b/gif/no_padding_no_strides_transposed.gif
diff --git a/gif/no_padding_strides.gif b/gif/no_padding_strides.gif
diff --git a/gif/padding_strides.gif b/gif/padding_strides.gif
diff --git a/gif/padding_strides_odd.gif b/gif/padding_strides_odd.gif
diff --git a/gif/padding_strides_odd_transposed.gif b/gif/padding_strides_odd_transposed.gif
diff --git a/gif/padding_strides_transposed.gif b/gif/padding_strides_transposed.gif
diff --git a/gif/same_padding_no_strides.gif b/gif/same_padding_no_strides.gif
diff --git a/pdf/arbitrary_padding_no_strides_transposed_00.pdf b/pdf/arbitrary_padding_no_strides_transposed_00.pdf
diff --git a/pdf/arbitrary_padding_no_strides_transposed_01.pdf b/pdf/arbitrary_padding_no_strides_transposed_01.pdf
diff --git a/pdf/arbitrary_padding_no_strides_transposed_02.pdf b/pdf/arbitrary_padding_no_strides_transposed_02.pdf
diff --git a/pdf/arbitrary_padding_no_strides_transposed_03.pdf b/pdf/arbitrary_padding_no_strides_transposed_03.pdf
diff --git a/pdf/full_padding_no_strides_transposed_00.pdf b/pdf/full_padding_no_strides_transposed_00.pdf
diff --git a/pdf/full_padding_no_strides_transposed_01.pdf b/pdf/full_padding_no_strides_transposed_01.pdf
diff --git a/pdf/full_padding_no_strides_transposed_02.pdf b/pdf/full_padding_no_strides_transposed_02.pdf
diff --git a/pdf/full_padding_no_strides_transposed_03.pdf b/pdf/full_padding_no_strides_transposed_03.pdf
diff --git a/pdf/no_padding_no_strides_00.pdf b/pdf/no_padding_no_strides_00.pdf
diff --git a/pdf/no_padding_no_strides_01.pdf b/pdf/no_padding_no_strides_01.pdf
diff --git a/pdf/no_padding_no_strides_02.pdf b/pdf/no_padding_no_strides_02.pdf
diff --git a/pdf/no_padding_no_strides_03.pdf b/pdf/no_padding_no_strides_03.pdf
diff --git a/pdf/no_padding_no_strides_transposed_00.pdf b/pdf/no_padding_no_strides_transposed_00.pdf
diff --git a/pdf/no_padding_no_strides_transposed_01.pdf b/pdf/no_padding_no_strides_transposed_01.pdf
diff --git a/pdf/no_padding_no_strides_transposed_02.pdf b/pdf/no_padding_no_strides_transposed_02.pdf
diff --git a/pdf/no_padding_no_strides_transposed_03.pdf b/pdf/no_padding_no_strides_transposed_03.pdf
diff --git a/pdf/no_padding_strides_00.pdf b/pdf/no_padding_strides_00.pdf
diff --git a/pdf/no_padding_strides_01.pdf b/pdf/no_padding_strides_01.pdf
diff --git a/pdf/no_padding_strides_02.pdf b/pdf/no_padding_strides_02.pdf
diff --git a/pdf/no_padding_strides_03.pdf b/pdf/no_padding_strides_03.pdf
diff --git a/pdf/numerical_average_pooling_00.pdf b/pdf/numerical_average_pooling_00.pdf
diff --git a/pdf/numerical_average_pooling_01.pdf b/pdf/numerical_average_pooling_01.pdf
diff --git a/pdf/numerical_average_pooling_02.pdf b/pdf/numerical_average_pooling_02.pdf
diff --git a/pdf/numerical_average_pooling_03.pdf b/pdf/numerical_average_pooling_03.pdf
diff --git a/pdf/numerical_average_pooling_04.pdf b/pdf/numerical_average_pooling_04.pdf
diff --git a/pdf/numerical_average_pooling_05.pdf b/pdf/numerical_average_pooling_05.pdf
diff --git a/pdf/numerical_average_pooling_06.pdf b/pdf/numerical_average_pooling_06.pdf
diff --git a/pdf/numerical_average_pooling_07.pdf b/pdf/numerical_average_pooling_07.pdf
diff --git a/pdf/numerical_average_pooling_08.pdf b/pdf/numerical_average_pooling_08.pdf
diff --git a/pdf/numerical_max_pooling_00.pdf b/pdf/numerical_max_pooling_00.pdf
diff --git a/pdf/numerical_max_pooling_01.pdf b/pdf/numerical_max_pooling_01.pdf
diff --git a/pdf/numerical_max_pooling_02.pdf b/pdf/numerical_max_pooling_02.pdf
diff --git a/pdf/numerical_max_pooling_03.pdf b/pdf/numerical_max_pooling_03.pdf
diff --git a/pdf/numerical_max_pooling_04.pdf b/pdf/numerical_max_pooling_04.pdf
diff --git a/pdf/numerical_max_pooling_05.pdf b/pdf/numerical_max_pooling_05.pdf
diff --git a/pdf/numerical_max_pooling_06.pdf b/pdf/numerical_max_pooling_06.pdf
diff --git a/pdf/numerical_max_pooling_07.pdf b/pdf/numerical_max_pooling_07.pdf
diff --git a/pdf/numerical_max_pooling_08.pdf b/pdf/numerical_max_pooling_08.pdf
diff --git a/pdf/numerical_no_padding_no_strides_00.pdf b/pdf/numerical_no_padding_no_strides_00.pdf
diff --git a/pdf/numerical_no_padding_no_strides_01.pdf b/pdf/numerical_no_padding_no_strides_01.pdf
diff --git a/pdf/numerical_no_padding_no_strides_02.pdf b/pdf/numerical_no_padding_no_strides_02.pdf
diff --git a/pdf/numerical_no_padding_no_strides_03.pdf b/pdf/numerical_no_padding_no_strides_03.pdf
diff --git a/pdf/numerical_no_padding_no_strides_04.pdf b/pdf/numerical_no_padding_no_strides_04.pdf
diff --git a/pdf/numerical_no_padding_no_strides_05.pdf b/pdf/numerical_no_padding_no_strides_05.pdf
diff --git a/pdf/numerical_no_padding_no_strides_06.pdf b/pdf/numerical_no_padding_no_strides_06.pdf
diff --git a/pdf/numerical_no_padding_no_strides_07.pdf b/pdf/numerical_no_padding_no_strides_07.pdf
diff --git a/pdf/numerical_no_padding_no_strides_08.pdf b/pdf/numerical_no_padding_no_strides_08.pdf
diff --git a/pdf/numerical_padding_strides_00.pdf b/pdf/numerical_padding_strides_00.pdf
diff --git a/pdf/numerical_padding_strides_01.pdf b/pdf/numerical_padding_strides_01.pdf
diff --git a/pdf/numerical_padding_strides_02.pdf b/pdf/numerical_padding_strides_02.pdf
diff --git a/pdf/numerical_padding_strides_03.pdf b/pdf/numerical_padding_strides_03.pdf
diff --git a/pdf/numerical_padding_strides_04.pdf b/pdf/numerical_padding_strides_04.pdf
diff --git a/pdf/numerical_padding_strides_05.pdf b/pdf/numerical_padding_strides_05.pdf
diff --git a/pdf/numerical_padding_strides_06.pdf b/pdf/numerical_padding_strides_06.pdf
diff --git a/pdf/numerical_padding_strides_07.pdf b/pdf/numerical_padding_strides_07.pdf
diff --git a/pdf/numerical_padding_strides_08.pdf b/pdf/numerical_padding_strides_08.pdf
diff --git a/pdf/padding_strides_00.pdf b/pdf/padding_strides_00.pdf
diff --git a/pdf/padding_strides_01.pdf b/pdf/padding_strides_01.pdf
diff --git a/pdf/padding_strides_02.pdf b/pdf/padding_strides_02.pdf
diff --git a/pdf/padding_strides_03.pdf b/pdf/padding_strides_03.pdf
diff --git a/pdf/padding_strides_odd_00.pdf b/pdf/padding_strides_odd_00.pdf
diff --git a/pdf/padding_strides_odd_01.pdf b/pdf/padding_strides_odd_01.pdf
diff --git a/pdf/padding_strides_odd_02.pdf b/pdf/padding_strides_odd_02.pdf
diff --git a/pdf/padding_strides_odd_03.pdf b/pdf/padding_strides_odd_03.pdf
diff --git a/pdf/padding_strides_odd_transposed_00.pdf b/pdf/padding_strides_odd_transposed_00.pdf
diff --git a/pdf/padding_strides_odd_transposed_01.pdf b/pdf/padding_strides_odd_transposed_01.pdf
diff --git a/pdf/padding_strides_odd_transposed_02.pdf b/pdf/padding_strides_odd_transposed_02.pdf
diff --git a/pdf/padding_strides_odd_transposed_03.pdf b/pdf/padding_strides_odd_transposed_03.pdf
diff --git a/pdf/padding_strides_transposed_00.pdf b/pdf/padding_strides_transposed_00.pdf
diff --git a/pdf/padding_strides_transposed_01.pdf b/pdf/padding_strides_transposed_01.pdf
diff --git a/pdf/padding_strides_transposed_02.pdf b/pdf/padding_strides_transposed_02.pdf
diff --git a/pdf/padding_strides_transposed_03.pdf b/pdf/padding_strides_transposed_03.pdf
diff --git a/pdf/same_padding_no_strides_00.pdf b/pdf/same_padding_no_strides_00.pdf
diff --git a/pdf/same_padding_no_strides_01.pdf b/pdf/same_padding_no_strides_01.pdf
diff --git a/pdf/same_padding_no_strides_02.pdf b/pdf/same_padding_no_strides_02.pdf
diff --git a/pdf/same_padding_no_strides_03.pdf b/pdf/same_padding_no_strides_03.pdf
diff --git a/templates/arithmetic_figure.txt b/templates/arithmetic_figure.txt
@@ -13,14 +13,19 @@
                     yslant=0.5,xslant=-0.7]
         \draw[step=10mm, base03, dashed, thick] (0,0) grid ({PADDING_TO});
         {INPUT_UNITS}
-        \draw[fill=base02, opacity=0.4] ({INPUT_GRID_FROM}) rectangle
-                                        ({INPUT_GRID_TO});
-        \draw[step=10mm, base03, thick] ({INPUT_GRID_FROM}) grid
-                                        ({INPUT_GRID_TO});
-        \coordinate (BL) at ({INPUT_BOTTOM_LEFT});
-        \coordinate (BR) at ({INPUT_BOTTOM_RIGHT});
-        \coordinate (TL) at ({INPUT_TOP_LEFT});
-        \coordinate (TR) at ({INPUT_TOP_RIGHT});
+
+        \foreach \x in {{ {INPUT_GRID_FROM_X},\number\numexpr {INPUT_GRID_FROM_X}+{DILATION},...,\number\numexpr {INPUT_GRID_TO_X}-1 }} {{
+            \foreach \y in {{ {INPUT_GRID_FROM_Y},\number\numexpr {INPUT_GRID_FROM_Y}+{DILATION},...,\number\numexpr {INPUT_GRID_TO_Y}-1 }} {{
+                \draw[fill=base02, opacity=0.4] (\x,\y) rectangle
+                                        (\x+1,\y+1);
+            }}
+        }}
+        \draw[step=10mm, base03, thick] ({INPUT_GRID_FROM_X}, {INPUT_GRID_FROM_Y}) grid
+                                        ({INPUT_GRID_TO_X}, {INPUT_GRID_TO_Y});
+        \coordinate (BL) at ({INPUT_GRID_FROM_X},{INPUT_GRID_FROM_Y});
+        \coordinate (BR) at ({INPUT_GRID_TO_X},{INPUT_GRID_FROM_Y});
+        \coordinate (TL) at ({INPUT_GRID_FROM_X},{INPUT_GRID_TO_Y});
+        \coordinate (TR) at ({INPUT_GRID_TO_X},{INPUT_GRID_TO_Y});
     \end{{scope}}
     \begin{{scope}}[xshift=-5, yshift={OUTPUT_ELEVATION},
                     every node/.append style={{yslant=0.5,xslant=-0.7}},