merge

_test_elemwise.py

@@ -135,7 +135,7 @@ class _test_CAReduce(unittest.TestCase):
                            ((2, 3, 4, 5), (0, 1, 3)),
                            ((), ())]:
             x = modes.build(Tensor('float64', [1 * (entry == 1) for entry in xsh], name = 'x'))
-            e = CAReduce(Add, [x], dimensions_to_reduce = tosum).out
+            e = CAReduce(Add, [x], axis = tosum).out
             f = linker(env([x], [e])).make_function(inplace = False)
             xv = numpy.asarray(numpy.random.rand(*xsh))
             zv = xv

elemwise.py

@@ -402,7 +402,7 @@ def make_broadcast(scalar_opclass, inplace_pattern = {}, name = None):
     scalar_name = scalar_opclass.__name__
     previous_doc = Broadcast.__doc__
 
-    scalar_doc = scalar_opclass.__doc__
+    scalar_doc = scalar_opclass.__doc__ or ""
     if scalar_doc:
         scalar_doc = """
     %(scalar_name)s documentation:
@@ -411,7 +411,7 @@ def make_broadcast(scalar_opclass, inplace_pattern = {}, name = None):
 
     doc = """
     Usage: %(name)s(*inputs)
-    Equivalent to: Broadcast(%(scalar_name)s, inputs, %(inplace_pattern)s)
+    Equivalent to: Broadcast(scalar.%(scalar_name)s, inputs, %(inplace_pattern)s)
 
     Performs Scalar %(scalar_name)s on each element of the
     input tensors.
@@ -460,15 +460,16 @@ def wrap_broadcast(op):
 
 class CAReduce(Op):
     """
-    Usage: CAReduce(scalar_opclass, inputs, dimensions_to_reduce = None)
+    Usage: CAReduce(scalar_opclass, inputs, axis = None)
 
     * scalar_opclass: a binary scalar op with only one output.
                       It will be instantiated as such:
                       scalar_opclass.__init__([Scalar(t.dtype) for t in inputs])
                       It must be commutative and associative.
     * inputs: list of Tensor instances
-    * dimensions_to_reduce: list of dimensions that we want to reduce
-                            if None, all dimensions are reduced
+    * axis: - the dimension along which we want to reduce
+            - list of dimensions that we want to reduce
+            - if None, all dimensions are reduced
 
     The output will have the same shape as the input minus the reduced
     dimensions. It will contain the result of accumulating all values
@@ -489,7 +490,7 @@ class CAReduce(Op):
     or/and/xor - but not subtract, divide or power).
     """
     
-    def __init__(self, scalar_opclass, inputs, dimensions_to_reduce = None):
+    def __init__(self, scalar_opclass, inputs, axis = None):
         inputs = map(astensor, inputs)
 
         self.shadow = scalar_opclass(*[Scalar(dtype = inputs[0].dtype) for i in xrange(len(inputs) + 1)])
@@ -498,32 +499,34 @@ class CAReduce(Op):
             raise NotImplementedError("CAReduce only supports binary functions with a single output.")
         if len(inputs) != 1:
             raise TypeError("Only one argument expected.")
-        if dimensions_to_reduce is None:
-            dimensions_to_reduce = range(len(inputs[0].broadcastable))
+        if axis is None:
+            axis = range(len(inputs[0].broadcastable))
+        elif isinstance(axis, int):
+            axis = [axis]
 
         self.inputs = inputs
         self.outputs = [Tensor(dtype = inputs[0].dtype,
-                               broadcastable = [x for i, x in enumerate(inputs[0].broadcastable) if i not in dimensions_to_reduce])]
+                               broadcastable = [x for i, x in enumerate(inputs[0].broadcastable) if i not in axis])]
 
-        self.dimensions_to_reduce = dimensions_to_reduce
+        self.axis = axis
         self.scalar_opclass = scalar_opclass
         self.ufunc = numpy.frompyfunc(self.shadow.impl, self.shadow.nin, self.shadow.nout)
 
     def desc(self):
-        return (self.__class__, self.scalar_opclass, tuple(self.dimensions_to_reduce))
+        return (self.__class__, self.scalar_opclass, tuple(self.axis))
         
     def strdesc(self):
-        if set(self.dimensions_to_reduce) != set(xrange(len(self.inputs[0].broadcastable))):
-            return "Reduce{%s}{%s}" % (self.scalar_opclass.__name__, "".join(str(x) for x in self.dimensions_to_reduce))
+        if set(self.axis) != set(xrange(len(self.inputs[0].broadcastable))):
+            return "Reduce{%s}{%s}" % (self.scalar_opclass.__name__, "".join(str(x) for x in self.axis))
         else:
             return "Reduce{%s}" % self.scalar_opclass.__name__
         
     def clone_with_new_inputs(self, *new_inputs):
-        return CAReduce(self.scalar_opclass, new_inputs, self.dimensions_to_reduce)
+        return CAReduce(self.scalar_opclass, new_inputs, self.axis)
         
     def perform(self):
         result = self.inputs[0].data
-        to_reduce = reversed(sorted(self.dimensions_to_reduce))
+        to_reduce = reversed(sorted(self.axis))
         if to_reduce:
             for dimension in to_reduce:
                 result = self.ufunc.reduce(result, dimension)
@@ -542,7 +545,7 @@ class CAReduce(Op):
         idtype = input.dtype_specs()[1]
         odtype = output.dtype_specs()[1]
 
-        tosum = self.dimensions_to_reduce
+        tosum = self.axis
 
         if tosum == ():
             return Broadcast(scalar.Identity, (input, ))._c_all(inames, onames, sub)
@@ -604,7 +607,7 @@ class CAReduce(Op):
 
     def __str__(self):
         input = self.inputs[0]
-        if len(input.broadcastable) == len(self.dimensions_to_reduce):
+        if len(input.broadcastable) == len(self.axis):
             return "%s:%s(%s)" % (self.__class__.__name__,
                                   self.scalar_opclass.__name__,
                                   str(input))
@@ -612,7 +615,7 @@ class CAReduce(Op):
             return "%s:%s(%s, axis = %s)" % (self.__class__.__name__,
                                              self.scalar_opclass.__name__,
                                              str(input),
-                                             self.dimensions_to_reduce)
+                                             self.axis)
         
         
 
@@ -645,16 +648,16 @@ def make_reduce(scalar_opclass, name = None):
         def __init__(self, *inputs, **kwargs):
             reducer.__init__(self, scalar_opclass, inputs, kwargs.get('axis', None))
         def clone_with_new_inputs(self, *new_inputs):
-            return New(*new_inputs, **dict(axis = self.dimensions_to_reduce))
+            return New(*new_inputs, **dict(axis = self.axis))
         def __str__(self):
             input = self.inputs[0]
-            if len(input.broadcastable) == len(self.dimensions_to_reduce):
+            if len(input.broadcastable) == len(self.axis):
                 return "%s(%s)" % (self.__class__.__name__,
                                    str(input))
             else:
                 return "%s(%s, axis = %s)" % (self.__class__.__name__,
                                               str(input),
-                                              self.dimensions_to_reduce)
+                                              self.axis)
     New.__name__ = name
     return New
 
@@ -662,12 +665,12 @@ _Sum = make_reduce(scalar.Add, '_Sum')
 class Sum(_Sum):
     __doc__ = _Sum.__doc__
     def grad(self, (x, ), (gz, )):
-        if self.dimensions_to_reduce == ():
+        if self.axis == ():
             return gz,
         new_dims = []
         i = 0
         for j, _ in enumerate(x.broadcastable):
-            if j in self.dimensions_to_reduce:
+            if j in self.axis:
                 new_dims.append('x')
             else:
                 new_dims.append(i)
@@ -681,7 +684,7 @@ def reduce(op):
     else:
         raise NotImplementedError("The scalar op class to reduce must be commutative and associative.")
     def instantiate(*inputs):
-        return reducer(op, inputs, dimensions_to_reduce)
+        return reducer(op, inputs, axis)
     return instantiate
 
 

gof/cc.py

@@ -421,7 +421,7 @@ class CLinker(Linker):
                 # orphans are not inputs so we'll just get fetch them when we initialize the struct and assume they stay the same
                 policy = [[get_c_declare, get_c_extract, get_c_cleanup],
                           [get_nothing, get_nothing, get_nothing]]
-            elif result in self.temps or not reuse_storage:
+            elif result in self.temps:
                 # temps don't need to be extracted from Python, so we call c_init rather than c_extract
                 # they do not need to be relayed to Python, so we don't sync
                 if result.c_is_simple() or not reuse_storage:
@@ -441,6 +441,8 @@ class CLinker(Linker):
                     # it is useful for complex outputs to reuse storage at each run, so we only clean up in the destructor
                     policy = [[get_c_declare, get_c_init, get_c_cleanup],
                               [get_nothing, get_nothing, get_c_sync]]
+            else:
+                raise Exception("what the fuck")
 
             builder, block = struct_result_codeblocks(result, policy, id, symbol, sub)
 

gof/env.py

@@ -155,12 +155,6 @@ class Env(graph.Graph):
         """
         if feature_class in self._features:
             return # the feature is already present
-        else:
-            for other_feature_class in self._features:
-                if issubclass(other_feature_class, feature_class):
-                    return
-                elif issubclass(feature_class, other_feature_class):
-                    self.__del_feature__(other_feature_class)
         self.__add_feature__(feature_class, do_import)
 
     def __add_feature__(self, feature_class, do_import):
@@ -193,14 +187,7 @@ class Env(graph.Graph):
                 pass
 
     def get_feature(self, feature_class):
-        try:
-            return self._features[feature_class]
-        except KeyError:
-            for other_feature_class in self._features:
-                if issubclass(other_feature_class, feature_class):
-                    return self._features[other_feature_class]
-            else:
-                raise
+        return self._features[feature_class]
 
     def has_feature(self, feature_class):
         try:
@@ -213,6 +200,18 @@ class Env(graph.Graph):
         "Same as len(self.clients(r))."
         return len(self.clients(r))
 
+    def edge(self, r):
+        return r in self.inputs or r in self.orphans()
+
+    def follow(self, r):
+        op = r.owner
+        if self.edge(r):
+            return None
+        else:
+            if op is None:
+                raise Exception("what the fuck")
+            return op.inputs
+    
     def ops(self):
         "All ops within the subgraph bound by env.inputs and env.outputs."
         return self._ops

opt.py

@@ -67,11 +67,9 @@ class DimShuffleLifter(opt.Optimizer):
             if r in seen:
                 return
             seen.add(r)
-            op = r.owner
-            if op is None \
-                    or op in env.inputs \
-                    or op in env.orphans():
+            if env.edge(r):
                 return
+            op = r.owner
             if isinstance(op, DimShuffle):
                 in_op = op.inputs[0].owner
                 if isinstance(in_op, DimShuffle):
@@ -121,9 +119,7 @@ def find_cliques(env, through_broadcast = False):
         # is False) a Result which needs to be broadcasted.
         
         op = r.owner
-        if r in env.inputs \
-                or r in env.orphans() \
-                or op is None \
+        if env.edge(r) \
                 or not isinstance(op, Broadcast) \
                 or len(op.outputs) > 1:
             # todo: handle multiple-output broadcast ops
@@ -155,7 +151,7 @@ def find_cliques(env, through_broadcast = False):
     cliques = []
 
     def find_cliques_helper(r):
-        if r in env.inputs or r in env.orphans():
+        if env.edge(r):
             return
         clique_inputs = seek_from(r)
         if clique_inputs is None:
@@ -218,7 +214,7 @@ class CliqueOptimizer(opt.Optimizer):
             if r in equiv:
                 return equiv[r]
             op = r.owner
-            if r in env.inputs or r in env.orphans():
+            if env.edge(r):
                 # For each leave we make a Scalar of the corresponding dtype
                 s = scalar.Scalar(dtype = r.dtype)
                 _r = r

scalar_opt.py

@@ -71,17 +71,22 @@ class Canonizer(gof.Optimizer):
 
         def canonize(r):
             
-            if r in env.inputs or r in env.orphans():
+#             if r in env.inputs or r in env.orphans():
+#                 return
+            next = env.follow(r)
+            if next is None:
                 return
             
             def flatten(r, nclients_check = True):
                 # Collapses a tree of main/inverse/reciprocal Ops (aka Mul/Div/Inv or Add/Sub/Neg)
                 # into a list of numerators and a list of denominators
                 # e.g. (x*(1/y))*(x/(z/a)) aka Mul(Mul(x, (Inv, y)), Div(x, Div(z, a))) -> [x, x, a], [z, y]
-                
-                op = r.owner
-                if op is None or r in env.inputs or r in env.orphans():
+
+                if env.edge(r):
                     return [r], []
+                op = r.owner
+#                 if op is None or r in env.inputs or r in env.orphans():
+#                     return [r], []
                 
                 results = [r2.dtype == r.dtype and flatten(r2) or ([r2], []) for r2 in op.inputs]
                 if isinstance(op, self.main) and (not nclients_check or env.nclients(r) == 1):
@@ -103,12 +108,15 @@ class Canonizer(gof.Optimizer):
             num, denum = flatten(r, False)
 
             if (num, denum) == ([r], []):
-                if r.owner is None:
-                    return
-                else:
-                    for input in r.owner.inputs:
-                        canonize(input)
-                    return
+                for input in (env.follow(r) or []):
+                    canonize(input)
+                return
+#                 if r.owner is None:
+#                     return
+#                 else:
+#                     for input in r.owner.inputs:
+#                         canonize(input)
+#                     return
 
             # Terms that are both in the num and denum lists cancel each other
             for d in list(denum):
@@ -194,7 +202,7 @@ def group_powers(env, num, denum):
         # and does d[base].append(power).
         for factor in list(seq):
             op = factor.owner
-            if op is None or factor in env.inputs or factor in env.orphans():
+            if env.edge(factor):
                 continue
             if isinstance(op, Exp):
                 d.setdefault('e', []).append(op.inputs[0])

tensor.py

@@ -14,6 +14,8 @@ import blas # for gemm, dot
 import elemwise as s2t
 import scalar as scal
 
+from functools import partial
+
 
 class Tensor(BaseTensor):
     """
@@ -100,6 +102,49 @@ def astensor(data, broadcastable=None, name=None):
 s2t.astensor = astensor
 
 
+# Easy constructors
+
+def _multi(*fns):
+    def f2(f, names):
+        if len(names) == 1:
+            return f(names)
+        else:
+            return [f(name) for name in names]
+    if len(fns) == 1:
+        return partial(f2, fns)
+    else:
+        return [partial(f2, f) for f in fns]
+
+def _int_float(f):
+    return partial(f, dtype = 'int64'), partial(f, dtype = 'float64') 
+
+def scalar(name, dtype = 'float64'):
+    return Tensor(name = name, dtype = dtype, broadcastable = ())
+iscalar, fscalar = _int_float(scalar)
+scalars, iscalars, fscalars = _multi(scalar, iscalar, fscalar)
+
+def vector(name, dtype = 'float64'):
+    return Tensor(name = name, dtype = dtype, broadcastable = (False))
+ivector, fvector = _int_float(vector)
+vectors, ivectors, fvectors = _multi(vector, ivector, fvector)
+
+def matrix(name, dtype = 'float64'):
+    return Tensor(name = name, dtype = dtype, broadcastable = (False, False))
+imatrix, fmatrix = _int_float(matrix)
+matrices, imatrices, fmatrices = _multi(matrix, imatrix, fmatrix)
+
+def row(name, dtype = 'float64'):
+    return Tensor(name = name, dtype = dtype, broadcastable = (True, False))
+irow, frow = _int_float(row)
+rows, irows, frows = _multi(row, irow, frow)
+
+def col(name, dtype = 'float64'):
+    return Tensor(name = name, dtype = dtype, broadcastable = (False, True))
+icol, fcol = _int_float(col)
+cols, icols, fcols = _multi(col, icol, fcol)
+
+
+
 ############################
 # Supporting Ops
 ############################