Check shape of cuDNN convolution result.

theano/gpuarray/dnn.py

@@ -24,7 +24,8 @@ from theano.tensor.nnet.abstract_conv import (AbstractConv2d,
                                               AbstractConv3d,
                                               AbstractConv3d_gradWeights,
                                               AbstractConv3d_gradInputs,
-                                              get_conv_output_shape)
+                                              get_conv_output_shape,
+                                              assert_conv_shape)
 from theano.tensor.signal.pool import (
     Pool, MaxPoolGrad, AveragePoolGrad)
 from . import pygpu
@@ -978,11 +979,12 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
             # that would be flipped by conv_mode='conv' in GpuDnnConvGradW.
             kerns = kerns[:, :, ::-1, ::-1]
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3))
-        shape2 = shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1
-        shape3 = shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1
-        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(
-            shape_i(kerns, 1, fgraph),
-            shape_i(img, 1, fgraph), shape2, shape3)
+        out_shp = (shape_i(kerns, 1, fgraph),
+                   shape_i(img, 1, fgraph),
+                   shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1,
+                   shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1)
+        out_shp = assert_conv_shape(out_shp)
+        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(*out_shp)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1),
                               conv_mode='cross', precision=precision)(out.shape)
         conv = gpu_dnn_conv_gradW()(img, kerns, out, desc)
@@ -996,11 +998,12 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
         img = gpu_contiguous(img)  # cudnn v2 rc3 need contiguous data
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3))
         conv_mode = 'cross' if conv_mode == 'conv' else 'conv'
-        shape2 = shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1
-        shape3 = shape_i(img, 3, fgraph) + shape_i(kerns, 3, fgraph) - 1
-        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(shape_i(img, 0, fgraph),
-                                                         shape_i(kerns, 1, fgraph),
-                                                         shape2, shape3)
+        out_shp = (shape_i(img, 0, fgraph),
+                   shape_i(kerns, 1, fgraph),
+                   shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1,
+                   shape_i(img, 3, fgraph) + shape_i(kerns, 3, fgraph) - 1)
+        out_shp = assert_conv_shape(out_shp)
+        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(*out_shp)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1),
                               conv_mode=conv_mode, precision=precision)(kerns.shape)
         return gpu_dnn_conv_gradI()(kerns, img, out, desc)
@@ -1020,6 +1023,7 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
     out_shp = get_conv_output_shape(ishape, kshape,
                                     desc_op.border_mode,
                                     desc_op.subsample)
+    out_shp = assert_conv_shape(out_shp)
     out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(*out_shp)
     return gpu_dnn_conv(algo=algo)(img, kerns, out, desc)
 
@@ -1092,12 +1096,13 @@ def dnn_conv3d(img, kerns, border_mode='valid', subsample=(1, 1, 1),
             # that would be flipped by conv_mode='conv' in GpuDnnConvGradW.
             kerns = kerns[:, :, ::-1, ::-1]
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3, 4))
-        shape2 = shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1
-        shape3 = shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1
-        shape4 = shape_i(img, 4, fgraph) - shape_i(kerns, 4, fgraph) + 1
-        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(
-            shape_i(kerns, 1, fgraph),
-            shape_i(img, 1, fgraph), shape2, shape3, shape4)
+        out_shp = (shape_i(kerns, 1, fgraph),
+                   shape_i(img, 1, fgraph),
+                   shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1,
+                   shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1,
+                   shape_i(img, 4, fgraph) - shape_i(kerns, 4, fgraph) + 1)
+        out_shp = assert_conv_shape(out_shp)
+        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(*out_shp)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1, 1),
                               conv_mode='cross', precision=precision)(out.shape)
         conv = gpu_dnn_conv_gradW()(img, kerns, out, desc)
@@ -1111,12 +1116,13 @@ def dnn_conv3d(img, kerns, border_mode='valid', subsample=(1, 1, 1),
         img = gpu_contiguous(img)  # cudnn v2 rc3 need contiguous data
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3, 4))
         conv_mode = 'cross' if conv_mode == 'conv' else 'conv'
-        shape2 = shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1
-        shape3 = shape_i(img, 3, fgraph) + shape_i(kerns, 3, fgraph) - 1
-        shape4 = shape_i(img, 4, fgraph) + shape_i(kerns, 4, fgraph) - 1
-        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(shape_i(img, 0, fgraph),
-                                                         shape_i(kerns, 1, fgraph),
-                                                         shape2, shape3, shape4)
+        out_shp = (shape_i(img, 0, fgraph),
+                   shape_i(kerns, 1, fgraph),
+                   shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1,
+                   shape_i(img, 3, fgraph) + shape_i(kerns, 3, fgraph) - 1,
+                   shape_i(img, 4, fgraph) + shape_i(kerns, 4, fgraph) - 1)
+        out_shp = assert_conv_shape(out_shp)
+        out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(*out_shp)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1, 1),
                               conv_mode=conv_mode, precision=precision)(kerns.shape)
         return gpu_dnn_conv_gradI()(kerns, img, out, desc)
@@ -1136,6 +1142,7 @@ def dnn_conv3d(img, kerns, border_mode='valid', subsample=(1, 1, 1),
     out_shp = get_conv_output_shape(ishape, kshape,
                                     desc_op.border_mode,
                                     desc_op.subsample)
+    out_shp = assert_conv_shape(out_shp)
     out = gpu_alloc_empty(ctx_name, dtype=img.dtype)(*out_shp)
     return gpu_dnn_conv(algo=algo)(img, kerns, out, desc)
 

theano/sandbox/cuda/dnn.py

@@ -14,7 +14,8 @@ from theano.gof.type import CDataType
 from theano.compile import optdb
 from theano.compile.ops import shape_i
 from theano.tensor.nnet import LogSoftmax, SoftmaxGrad
-from theano.tensor.nnet.abstract_conv import get_conv_output_shape
+from theano.tensor.nnet.abstract_conv import (get_conv_output_shape,
+                                              assert_conv_shape)
 from theano.tensor.signal.pool import (
     Pool, MaxPoolGrad, AveragePoolGrad)
 from theano.sandbox.cuda.type import CudaNdarrayType
@@ -1132,10 +1133,12 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
             # that would be flipped by conv_mode='conv' in GpuDnnConvGradW.
             kerns = kerns[:, :, ::-1, ::-1]
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3))
-        shape2 = shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1
-        shape3 = shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1
-        out = gpu_alloc_empty(shape_i(kerns, 1, fgraph),
-                              shape_i(img, 1, fgraph), shape2, shape3)
+        out_shp = (shape_i(kerns, 1, fgraph),
+                   shape_i(img, 1, fgraph),
+                   shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1,
+                   shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1)
+        out_shp = assert_conv_shape(out_shp)
+        out = gpu_alloc_empty(*out_shp)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1),
                               conv_mode='cross', precision=precision)(img.shape,
                                                                       out.shape)
@@ -1149,10 +1152,12 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
         img = gpu_contiguous(img)
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3))
         conv_mode = 'cross' if conv_mode == 'conv' else 'conv'
-        shape2 = shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1
-        shape3 = shape_i(img, 3, fgraph) + shape_i(kerns, 3, fgraph) - 1
-        out = gpu_alloc_empty(shape_i(img, 0, fgraph),
-                              shape_i(kerns, 1, fgraph), shape2, shape3)
+        out_shp = (shape_i(img, 0, fgraph),
+                   shape_i(kerns, 1, fgraph),
+                   shape_i(img, 2, fgraph) + shape_i(kerns, 2, fgraph) - 1,
+                   shape_i(img, 3, fgraph) + shape_i(kerns, 3, fgraph) - 1)
+        out_shp = assert_conv_shape(out_shp)
+        out = gpu_alloc_empty(*out_shp)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1),
                               conv_mode=conv_mode, precision=precision)(out.shape,
                                                                         kerns.shape)
@@ -1170,6 +1175,7 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
     out_shp = GpuDnnConv.get_out_shape(img.shape, kerns.shape,
                                        desc_op.border_mode,
                                        desc_op.subsample)
+    out_shp = assert_conv_shape(out_shp)
     out = gpu_alloc_empty(*out_shp)
     return GpuDnnConv(algo=algo)(img, kerns, out, desc)
 
@@ -1248,11 +1254,13 @@ def dnn_conv3d(img, kerns, border_mode='valid', subsample=(1, 1, 1),
             # that would be flipped by conv_mode='conv' in GpuDnnConvGradW.
             kerns = kerns[:, :, ::-1, ::-1, ::-1]
         kerns = gpu_contiguous(kerns.dimshuffle(1, 0, 2, 3, 4))
-        shape2 = shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1
-        shape3 = shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1
-        shape4 = shape_i(img, 4, fgraph) - shape_i(kerns, 4, fgraph) + 1
-        out = gpu_alloc_empty(shape_i(kerns, 1, fgraph),
-                              shape_i(img, 1, fgraph), shape2, shape3, shape4)
+        out_shp = (shape_i(kerns, 1, fgraph),
+                   shape_i(img, 1, fgraph),
+                   shape_i(img, 2, fgraph) - shape_i(kerns, 2, fgraph) + 1,
+                   shape_i(img, 3, fgraph) - shape_i(kerns, 3, fgraph) + 1,
+                   shape_i(img, 4, fgraph) - shape_i(kerns, 4, fgraph) + 1)
+        out_shp = assert_conv_shape(out_shp)
+        out = gpu_alloc_empty(*out_shp)
         desc = GpuDnnConvDesc(border_mode='valid', subsample=(1, 1, 1),
                               conv_mode='cross', precision=precision)(img.shape,
                                                                       out.shape)
@@ -1271,6 +1279,7 @@ def dnn_conv3d(img, kerns, border_mode='valid', subsample=(1, 1, 1),
     out_shp = GpuDnnConv3d.get_out_shape(img.shape, kerns.shape,
                                          desc_op.border_mode,
                                          desc_op.subsample)
+    out_shp = assert_conv_shape(out_shp)
     out = gpu_alloc_empty(*out_shp)
     return GpuDnnConv3d(algo=algo)(img, kerns, out, desc)
 

theano/tensor/nnet/abstract_conv.py

@@ -11,6 +11,7 @@ import theano
 
 from theano.tensor import as_tensor_variable, patternbroadcast
 from theano.tensor import get_scalar_constant_value, NotScalarConstantError
+from theano.tensor.opt import Assert
 from theano.gof import Apply, Op
 
 from six.moves import xrange
@@ -428,6 +429,34 @@ def check_conv_gradinputs_shape(image_shape, kernel_shape, output_shape,
                for (given, computed) in zip(output_shape, computed_output_shape))
 
 
+def assert_conv_shape(shape):
+    """This function adds Assert nodes that check if shape is a valid convolution shape.
+
+    Parameters
+    ----------
+    shape: tuple of int (symbolic or numeric) corresponding to the input, output or
+        kernel shape of a convolution. For input and output, the first elements should
+        should be the batch size and number of channels. For kernels, the first and
+        second elements should contain the number of input and output channels.
+        The remaining dimensions are the convolution dimensions.
+
+    Returns
+    -------
+    Returns a tuple similar to the given `shape`, but with each element wrapped in
+    an `Assert` op that checks that dimension. The first two dimensions should be
+    larger than or equal to zero. The convolution dimensions should be larger than zero.
+    """
+    out_shape = []
+    for i, n in enumerate(shape):
+        if i < 2:
+            assert_shp = Assert('The convolution would produce an invalid shape (dim[%d] < 0).' % i)
+            out_shape.append(assert_shp(n, theano.tensor.ge(n, 0)))
+        else:
+            assert_shp = Assert('The convolution would produce an invalid shape (dim[%d] <= 0).' % i)
+            out_shape.append(assert_shp(n, theano.tensor.gt(n, 0)))
+    return tuple(out_shape)
+
+
 def conv2d(input,
            filters,
            input_shape=None,

theano/tensor/nnet/tests/test_abstract_conv.py

@@ -13,7 +13,8 @@ from theano.tensor.nnet import corr, corr3d, abstract_conv as conv
 from theano.tensor.nnet.abstract_conv import (get_conv_output_shape,
                                               get_conv_gradweights_shape,
                                               get_conv_gradinputs_shape,
-                                              check_conv_gradinputs_shape)
+                                              check_conv_gradinputs_shape,
+                                              assert_conv_shape)
 from theano.tensor.nnet.abstract_conv import AbstractConv2d
 from theano.tensor.nnet.abstract_conv import AbstractConv2d_gradInputs
 from theano.tensor.nnet.abstract_conv import AbstractConv2d_gradWeights
@@ -211,6 +212,20 @@ class TestConvGradInputsShape(unittest.TestCase):
                         self.assertEqual(computed_kernel_shape, kernel_shape_with_None)
 
 
+class TestAssertConvShape(unittest.TestCase):
+    def test_basic(self):
+        shape = tuple(tensor.iscalar() for i in range(4))
+        f = theano.function(shape, assert_conv_shape(shape))
+
+        self.assertEqual([1, 2, 3, 4], f(1, 2, 3, 4))
+        self.assertEqual([0, 0, 1, 1], f(0, 0, 1, 1))
+        assert_raises(AssertionError, f, 3, 3, 3, 0)
+        assert_raises(AssertionError, f, 3, 3, 0, 3)
+        assert_raises(AssertionError, f, 3, 3, -1, 3)
+        assert_raises(AssertionError, f, 3, -1, 3, 3)
+        assert_raises(AssertionError, f, -1, 3, 3, 3)
+
+
 class BaseTestConv(object):
     def get_output_shape(self, inputs_shape, filters_shape,
                          subsample, border_mode, filter_dilation):