Merge pull request #756 from lamblin/fix_conv_subsampling

theano/tensor/nnet/Conv3D.py

@@ -80,10 +80,15 @@ class Conv3D(theano.Op):
         #quit(-1)
         #dCdH = printing.Print("dCdH = ",["shape"])
 
-        dCdV = ConvTransp3D.convTransp3D(W, T.zeros_like(V[0,0,0,0,:]), d, dCdH, V.shape[1:4] )
+        # Make sure the broadcasting pattern of the gradient is the the same
+        # as the initial variable
+        dCdV = ConvTransp3D.convTransp3D(W, T.zeros_like(V[0,0,0,0,:]), d, dCdH, V.shape[1:4])
+        dCdV = T.patternbroadcast(dCdV, V.broadcastable)
         WShape = W.shape
         dCdW = ConvGrad3D.convGrad3D(V,d,WShape,dCdH)
+        dCdW = T.patternbroadcast(dCdW, W.broadcastable)
         dCdb = T.sum(dCdH, axis=(0,1,2,3))
+        dCdb = T.patternbroadcast(dCdb, b.broadcastable)
         dCdd = None #not differentiable, since d is not continuous
 
         if 'name' in dir(dCdH) and dCdH.name is not None:

theano/tensor/nnet/conv.py

@@ -716,6 +716,26 @@ class ConvOp(Op):
         if self.imshp != self.imshp_logical or self.kshp != self.kshp_logical:
             raise NotImplementedError('todo')
 
+        if self.out_mode == 'valid' and (self.dx, self.dy) != (1, 1):
+            # Use the gradient as defined in conv3D, because the implementation
+            # by Conv is slow (about 3x slower than conv3D, and probably 10x
+            # slower than it could be), and incorrect when dx or dy > 2.
+
+            # build a "node", that should be equivalent to the one given by
+            # self.make_node, but using conv3D instead of self.
+            tmp_node = theano.tensor.nnet.conv3D(
+                    V=inputs.dimshuffle(0, 2, 3, 'x', 1),
+                    W=kerns[:, :, ::-1, ::-1].dimshuffle(0, 2, 3, 'x', 1),
+                    b=theano.tensor.alloc(numpy.asarray(0, dtype=kerns.dtype), kerns.shape[0]),
+                    d=(self.dx, self.dy, 1))
+            node = theano.tensor.addbroadcast(tmp_node, 3).dimshuffle(0, 4, 1, 2)
+
+            # mimic what happens inside theano.grad: get the input gradient
+            # of the final cost wrt all variables involved.
+            tmp_gmap = theano.gradient.grad_sources_inputs([(node, gz)], [inputs, kerns])
+
+            return [tmp_gmap[inputs], tmp_gmap[kerns]]
+
         if self.dx not in (1, 2) or self.dy not in (1, 2):
             raise NotImplementedError("ERROR: We disable ConvOp.grad now when dx or "\
                     "dy are different from 1 and 2, as there is a bug in it.")

theano/tensor/nnet/tests/test_conv.py

@@ -236,10 +236,11 @@ class TestConv2D(unittest.TestCase):
         self.validate((3, 2, 7, 5), (5, 2, 2, 3), 'valid', subsample=(2, 2))
         self.validate((3, 2, 7, 5), (5, 2, 2, 3), 'full', subsample=(2, 2))
         self.validate((3, 2, 7, 5), (5, 2, 2, 3), 'valid', subsample=(2, 1))
+        self.validate((1, 1, 6, 6), (1, 1, 3, 3), 'valid', subsample=(3, 3))
 
-        # Fails as of 2012-04-12
+        # Fails as of 2012-07-11
         self.assertRaises(NotImplementedError, self.validate, (1, 1, 6, 6),
-                          (1, 1, 3, 3), 'valid', subsample=(3, 3))
+                          (1, 1, 3, 3), 'full', subsample=(3, 3))
 
     def test_shape_Constant_tensor(self):
         """