Merge pull request #3682 from nouiz/abs_conv

doc/library/tensor/basic.txt

@@ -510,7 +510,7 @@ TensorVariable
     .. method:: diagonal(offset=0, axis1=0, axis2=1)
     .. method:: astype(dtype)
     .. method:: take(indices, axis=None, mode='raise')
-    .. method:: copy()
+    .. method:: copy() Return a new symbolic variable that is a copy of the variable. Does not copy the tag.
     .. method:: norm(L, axis=None)
     .. method:: nonzero(self, return_matrix=False)
     .. method:: nonzero_values(self)

theano/misc/check_blas.py

@@ -262,6 +262,7 @@ if __name__ == "__main__":
         GTX Titan Black   0.64s  0.64s
         GTX Titan(D15U-50)
         GTX 780
+        GTX 980 Ti        0.41s
         GTX 980
         GTX 970
         GTX 680                  1.57s

theano/sandbox/gpuarray/basic_ops.py

@@ -422,6 +422,7 @@ class GpuFromHost(Op):
           }
         } else {
           Py_XDECREF(%(out)s);
+          // This method will release the GIL when needed.
           %(out)s = pygpu_fromhostdata(PyArray_DATA(%(name)s_tmp),
                                        get_typecode((PyObject *)PyArray_DESCR(%(name)s_tmp)),
                                        PyArray_NDIM(%(name)s_tmp),

theano/tensor/nnet/abstract_conv.py

@@ -103,10 +103,10 @@ def conv2d(input,
            border_mode='valid',
            subsample=(1, 1),
            filter_flip=True):
-    """
-    This function will build the symbolic graph for convolving a mini-batch of a
-    stack of 2D inputs with a set of 2D filters. The implementation is modelled
-    after Convolutional Neural Networks (CNN).
+    """This function will build the symbolic graph for convolving a
+    mini-batch of a stack of 2D inputs with a set of 2D filters. The
+    implementation is modelled after Convolutional Neural Networks
+    (CNN).
 
     :type input: symbolic 4D tensor
     :param input: mini-batch of feature map stacks, of shape
@@ -153,11 +153,20 @@ def conv2d(input,
     :param filter_flip: If ``True``, will flip the filter rows and columns
         before sliding them over the input. This operation is normally referred
         to as a convolution, and this is the default. If ``False``, the filters
-        are not flipped and the operation is referred to as a cross-correlation.
+        are not flipped and the operation is referred to as a
+        cross-correlation.
 
     :rtype: symbolic 4D tensor
     :return: set of feature maps generated by convolutional layer. Tensor is
         of shape (batch size, output channels, output rows, output columns)
+
+    :note: If CuDNN is available, it will be used on the
+        GPU. Otherwise, it is the *CorrMM* convolution that will be used
+        "caffe style convolution".
+
+    :note: This is only supported in Theano 0.8 or the development
+        version until it is released.
+
     """
 
     conv_op = AbstractConv2d(imshp=input_shape,
@@ -169,9 +178,10 @@ def conv2d(input,
 
 
 class BaseAbstractConv2d(Op):
-    """
-    Base class for AbstractConv
-    Define an abstract convolution op that will be replaced with the appropriate implementation
+    """Base class for AbstractConv
+
+    Define an abstract convolution op that will be replaced with the
+    appropriate implementation
 
     :type imshp: None, tuple/list of len 4 of int or Constant variable
     :param imshp: The shape of the input parameter.
@@ -211,7 +221,9 @@ class BaseAbstractConv2d(Op):
     :param filter_flip: If ``True``, will flip the filter rows and columns
         before sliding them over the input. This operation is normally referred
         to as a convolution, and this is the default. If ``False``, the filters
-        are not flipped and the operation is referred to as a cross-correlation.
+        are not flipped and the operation is referred to as a
+        cross-correlation.
+
     """
     check_broadcast = False
     __props__ = ('border_mode', 'subsample', 'filter_flip', 'imshp', 'kshp')
@@ -270,7 +282,8 @@ class AbstractConv2d(BaseAbstractConv2d):
                  subsample=(1, 1),
                  filter_flip=True):
         super(AbstractConv2d, self).__init__(imshp, kshp,
-                                             border_mode, subsample, filter_flip)
+                                             border_mode, subsample,
+                                             filter_flip)
 
     def make_node(self, img, kern):
         if img.type.ndim != 4:
@@ -319,7 +332,9 @@ class AbstractConv2d_gradWeights(BaseAbstractConv2d):
                  subsample=(1, 1),
                  filter_flip=True):
         super(AbstractConv2d_gradWeights, self).__init__(imshp, kshp,
-                                                         border_mode, subsample, filter_flip)
+                                                         border_mode,
+                                                         subsample,
+                                                         filter_flip)
 
     # Update shape/height_width
     def make_node(self, img, topgrad, shape):
@@ -336,7 +351,8 @@ class AbstractConv2d_gradWeights(BaseAbstractConv2d):
         return Apply(self, [img, topgrad, shape], [output])
 
     def perform(self, node, inp, out_):
-        raise NotImplementedError('AbstractConv2d_gradWeight theano optimization failed')
+        raise NotImplementedError(
+            'AbstractConv2d_gradWeight theano optimization failed')
 
     def grad(self, inp, grads):
         bottom, top = inp[:2]
@@ -344,7 +360,10 @@ class AbstractConv2d_gradWeights(BaseAbstractConv2d):
         d_bottom = AbstractConv2d_gradInputs(self.imshp, self.kshp,
                                              self.border_mode,
                                              self.subsample,
-                                             self.filter_flip)(weights, top, bottom.shape[-2:])
+                                             self.filter_flip)(
+                                                 weights,
+                                                 top,
+                                                 bottom.shape[-2:])
         d_top = AbstractConv2d(self.imshp,
                                self.kshp,
                                self.border_mode,
@@ -373,7 +392,9 @@ class AbstractConv2d_gradInputs(BaseAbstractConv2d):
                  subsample=(1, 1),
                  filter_flip=True):
         super(AbstractConv2d_gradInputs, self).__init__(imshp, kshp,
-                                                        border_mode, subsample, filter_flip)
+                                                        border_mode,
+                                                        subsample,
+                                                        filter_flip)
 
     # Update shape/height_width
     def make_node(self, kern, topgrad, shape):
@@ -390,16 +411,20 @@ class AbstractConv2d_gradInputs(BaseAbstractConv2d):
         return Apply(self, [kern, topgrad, shape], [output])
 
     def perform(self, node, inp, out_):
-        raise NotImplementedError('AbstractConv2d_gradWeight theano optimization failed')
+        raise NotImplementedError(
+            'AbstractConv2d_gradWeight theano optimization failed')
 
     def grad(self, inp, grads):
         weights, top = inp[:2]
         bottom, = grads
         d_weights = AbstractConv2d_gradWeights(self.imshp, self.kshp,
                                                self.border_mode,
-                                               self.subsample)(bottom, top, weights.shape[-2:])
+                                               self.subsample)(
+                                                   bottom, top,
+                                                   weights.shape[-2:])
         d_top = AbstractConv2d(self.imshp, self.kshp,
-                               self.border_mode, self.subsample)(bottom, weights)
+                               self.border_mode, self.subsample)(
+                                   bottom, weights)
         d_height_width = (theano.gradient.DisconnectedType()(),)
         return (d_weights, d_top) + d_height_width
 

theano/tensor/nnet/conv.py

@@ -881,8 +881,9 @@ class ConvOp(OpenMPOp):
 
         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.")
+                "ERROR: We disable ConvOp.grad now when output_mode is not"
+                " 'valid' and dx or dy are greater than 2, as there is a bug"
+                " in it. See `abstract_conv2d <>`_ for a version that support this.")
 
         all_shape = self.has_all_shape(self.imshp, self.kshp,
                                        self.nkern, self.bsize)

theano/tensor/var.py

@@ -535,7 +535,10 @@ class _tensor_py_operators(object):
 
     # COPYING
     def copy(self, name=None):
-        """Copy a variable and optionally assign a name."""
+        """Return a symbolic copy and optionally assign a name.
+
+        Does not copy the tags.
+        """
         copied_variable = theano.tensor.basic.tensor_copy(self)
         copied_variable.name = name
         return copied_variable