Merge pull request #2854 from nouiz/fix_warning

theano/compile/debugmode.py

@@ -293,6 +293,10 @@ class BadOptimization(DebugModeError):
                 nv - ov))
             print >> ssio, "  Std Abs Diff: ", numpy.std(numpy.absolute(
                 nv - ov))
+            arg_max_val = numpy.argmax(numpy.absolute(nv - ov))
+            values_at_max = (nv.flatten()[arg_max_val],
+                             ov.flatten()[arg_max_val])
+            print >> ssio, "  Value at Max Diff: ", values_at_max
 
             # N.B. the maximum(..., 1e-8) protects against div by 0 when
             #      nv == ov == 0
@@ -304,6 +308,10 @@ class BadOptimization(DebugModeError):
             print >> ssio, "  Mean Rel Diff: ", numpy.mean(reldiff)
             print >> ssio, "  Median Rel Diff: ", numpy.median(reldiff)
             print >> ssio, "  Std Rel Diff: ", numpy.std(reldiff)
+            arg_max_val = numpy.argmax(reldiff)
+            values_at_max = (nv.flatten()[arg_max_val],
+                             ov.flatten()[arg_max_val])
+            print >> ssio, "  Value at Max Diff: ", values_at_max
             # only if all succeeds to we add anything to sio
             print >> sio, ssio.getvalue()
         except Exception:
@@ -1559,10 +1567,13 @@ class _VariableEquivalenceTracker(object):
         if append_reason:
             # N.B. compute the debugprint now, because future
             # optimizations will change the graph
+            done = dict()
             self.reasons[new_r].append((reason,
                 r,
-                debugprint(r, prefix='  ', depth=6, file=StringIO()).getvalue(),
-                debugprint(new_r, prefix='  ',  depth=6, file=StringIO()).getvalue()))
+                debugprint(r, prefix='  ', depth=6,
+                           file=StringIO(), done=done).getvalue(),
+                debugprint(new_r, prefix='  ',  depth=6,
+                           file=StringIO(), done=done).getvalue()))
             self.replaced_by[r].append((reason, new_r))
 
         if r in self.equiv:

theano/printing.py

@@ -34,7 +34,8 @@ VALID_ASSOC = set(['left', 'right', 'either'])
 
 
 def debugprint(obj, depth=-1, print_type=False,
-               file=None, ids='CHAR', stop_on_name=False):
+               file=None, ids='CHAR', stop_on_name=False,
+               done=None):
     """Print a computation graph as text to stdout or a file.
 
     :type obj: Variable, Apply, or Function instance
@@ -53,6 +54,9 @@ def debugprint(obj, depth=-1, print_type=False,
                 "" - don't print an identifier
     :param stop_on_name: When True, if a node in the graph has a name,
                          we don't print anything below it.
+    :type done: None or dict
+    :param done: A dict where we store the ids of printed node.
+        Useful to have multiple call to debugprint share the same ids.
 
     :returns: string if `file` == 'str', else file arg
 
@@ -80,6 +84,7 @@ def debugprint(obj, depth=-1, print_type=False,
         _file = sys.stdout
     else:
         _file = file
+    if done is None:
         done = dict()
     results_to_print = []
     profile_list = []

theano/sandbox/cuda/dnn.py

@@ -1718,17 +1718,19 @@ if True:
     @register_opt('cudnn')
     @local_optimizer([SoftmaxGrad])
     def local_softmax_dnn_grad(node):
-        if (
-            isinstance(node.op, SoftmaxGrad)
-            and (isinstance(node.inputs[0].owner.op, HostFromGpu)
-                 or isinstance(node.inputs[1].owner.op, HostFromGpu))
-        ):
+        if (isinstance(node.op, SoftmaxGrad) and
+            ((node.inputs[0].owner and
+              isinstance(node.inputs[0].owner.op, HostFromGpu))
+             or (node.inputs[1].owner and
+                 isinstance(node.inputs[1].owner.op, HostFromGpu)))):
             if not dnn_available():
                 return
             ins = []
             for n in node.inputs:
                 if isinstance(n.owner.op, HostFromGpu):
                     n = n.owner.inputs[0]
+                if n.ndim != 2:
+                    return
                 ins.append(n.dimshuffle(0, 1, 'x', 'x'))
 
             out = GpuDnnSoftmaxGrad(

theano/sandbox/cuda/tests/test_nnet.py

@@ -416,7 +416,7 @@ class test_SoftMax(unittest.TestCase):
         # Verify that the SoftmaxGrad -> GpuDnnSoftmaxGrad optimization is not
         # applied when cudnn is excluded or not available
         mode_wo_cudnn = mode_with_gpu.excluding("cudnn")
-        y = T.vector('y')
+        y = T.fvector('y')
         f = theano.function(
             [y],
             T.grad(T.nnet.softmax(y).mean(), y),
@@ -435,3 +435,22 @@ class test_SoftMax(unittest.TestCase):
                         i.op,
                         theano.tensor.nnet.SoftmaxGrad
                     )]) == 1)
+
+        # Verify that the SoftmaxGrad -> GpuDnnSoftmaxGrad do not
+        # crash with manual graph
+        y = T.fvector('y')
+        o = theano.tensor.nnet.SoftmaxGrad()(y, y*2)
+        f = theano.function([y], o, mode=mode_with_gpu)
+        sorted_f = f.maker.fgraph.toposort()
+        assert(len([i
+                    for i in sorted_f
+                    if isinstance(
+                        i.op,
+                        theano.sandbox.cuda.dnn.GpuDnnSoftmaxGrad
+                    )]) == 1)
+        assert(len([i
+                    for i in sorted_f
+                    if isinstance(
+                        i.op,
+                        theano.tensor.nnet.SoftmaxGrad
+                    )]) == 0)

theano/tensor/nnet/nnet.py

@@ -279,21 +279,19 @@ class SoftmaxGrad(gof.Op):
     nin = 2
     nout = 1
 
-    def __init__(self, **kwargs):
-        gof.Op.__init__(self, **kwargs)
-
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return tensor.hashtype(self)
-
-    def __str__(self):
-        return self.__class__.__name__
+    __props__ = ()
 
-    def make_node(self, dy, sm, **kwargs):
+    def make_node(self, dy, sm):
         dy = tensor.as_tensor_variable(dy)
         sm = tensor.as_tensor_variable(sm)
+        if dy.type.ndim not in (1, 2) \
+                or dy.type.dtype not in tensor.float_dtypes:
+            raise ValueError('dy must be 1-d or 2-d tensor of floats. Got ',
+                             dy.type)
+        if dy.ndim == 1:
+            dy = tensor.shape_padleft(dy, n_ones=1)
+        if sm.ndim == 1:
+            sm = tensor.shape_padleft(sm, n_ones=1)
         return Apply(self, [dy, sm], [sm.type.make_variable()])
 
     def perform(self, node, input_storage, output_storage):
@@ -394,24 +392,14 @@ class Softmax(gof.Op):
 
     nin = 1
     nout = 1
-
-    def __init__(self, **kwargs):
-        gof.Op.__init__(self, **kwargs)
-
-    def __eq__(self, other):
-        return type(self) == type(other)
-
-    def __hash__(self):
-        return hash(type(self))
-
-    def __str__(self):
-        return self.__class__.__name__
+    __props__ = ()
 
     def make_node(self, x):
         x = tensor.as_tensor_variable(x)
         if x.type.ndim not in (1, 2) \
                 or x.type.dtype not in tensor.float_dtypes:
-            raise ValueError('x must be 1-d or 2-d tensor of floats. Got ', x.type)
+            raise ValueError('x must be 1-d or 2-d tensor of floats. Got ',
+                             x.type)
         if x.ndim == 1:
             x = tensor.shape_padleft(x, n_ones=1)
         return Apply(self, [x], [x.type()])