Merge pull request #102 from lucasb-eyer/upsample-perforated

Implement "perforated" upsampling mode.

Author: lucasb-eyer

DeepFried2/layers/UpSample.py

@@ -2,7 +2,7 @@
 
 
 class UpSample(df.Module):
-    def __init__(self, upsample=(2,2), axes=[-2, -1], output_shape=None):
+    def __init__(self, upsample=(2,2), axes=[-2, -1], output_shape=None, mode='repeat'):
         """
         Upsamples an input (nearest-neighbour, repeat) `upsample` times along `axes`.
 
@@ -11,19 +11,35 @@ def __init__(self, upsample=(2,2), axes=[-2, -1], output_shape=None):
             If all `axes` should be upsampled by the same factor, `upsample` can also be just that factor.
         - `output_shape` can be used to crop the upsampled result to a desired shape, it follows `axes` just like `upsample` does.
             (TODO: move this part into a separate `Restrict` module)
+        - `mode` specifies how upsampling happens. Currently supported are:
+            - `repeat`: upsample by repeating values, aka "nearest" (the default).
+            - `perforated`: Put values in top-left corner, i.e. [1,2] becomes [1,0,2,0].
+                            This comes from http://www.brml.org/uploads/tx_sibibtex/281.pdf
         """
         df.Module.__init__(self)
         self.axes = axes
         self.upsample = df.utils.expand(upsample, len(axes), "upsample factor")
         self.output_shape = df.utils.expand(output_shape, len(axes), "output shape")
+        self.upsample_mode = mode
 
     def symb_forward(self, symb_input):
         """symb_input shape: 2D: (n_input, channels, height, width)
                              3D: (n_input, channels, depth, height, width)
         """
-        res = symb_input
-        for f, ax in zip(self.upsample, self.axes):
-            res = df.T.repeat(res, f, axis=ax)
+        if self.upsample_mode == 'repeat':
+            res = symb_input
+            for f, ax in zip(self.upsample, self.axes):
+                res = df.T.repeat(res, f, axis=ax)
+        elif self.upsample_mode == 'perforated':
+            shape = list(symb_input.shape)
+            slices = [slice(None)]*symb_input.ndim
+            for f, ax in zip(self.upsample, self.axes):
+                shape[ax] *= f
+                slices[ax] = slice(None, None, f)
+            res = df.T.zeros(shape, symb_input.dtype)
+            res = df.T.set_subtensor(res[tuple(slices)], symb_input)
+        else:
+            raise ValueError("Unsupported upsampling mode '{}'".format(self.upsample_mode))
 
         # TODO: move this out to its own `Restrict` module.
         if self.output_shape is not None: