Merge pull request #1797 from nouiz/fast_opt

theano/compile/function_module.py

@@ -1077,6 +1077,7 @@ class FunctionMaker(object):
         self.mode = mode
         self.accept_inplace = accept_inplace
         self.function_builder = function_builder
+        self.on_unused_input = on_unused_input  # Used only for the pickling
 
         self.required = [(i.value is None) for i in self.inputs]
         self.refeed = [
@@ -1215,6 +1216,7 @@ def _pickle_FunctionMaker(self):
                 accept_inplace=self.accept_inplace,
                 function_builder=self.function_builder,
                 profile=self.profile,
+                on_unused_input=self.on_unused_input,
                 )
     return (_constructor_FunctionMaker, (kwargs,))
 

theano/tensor/basic.py

@@ -508,13 +508,31 @@ class EmptyConstantError(NotScalarConstantError):
     """
 
 
+def numpy_scalar(data):
+    """ Return a scalar stored in a numpy ndarray, or raise
+    NotScalarConstantError if the numpy ndarray is not a scalar
+    """
+
+    # handle case where data is numpy.array([])
+    if data.ndim > 0 and  (len(data.shape) == 0 or
+        __builtins__['max'](data.shape) == 0):
+        assert numpy.all(numpy.array([]) == data)
+        raise EmptyConstantError()
+    try:
+        numpy.complex(data)  # works for all numeric scalars
+        return data
+    except Exception:
+        raise NotScalarConstantError(
+            'v.data is non-numeric, non-scalar, or has more than one'
+            ' unique value', data)
+
 get_scalar_constant_value_elemwises = (
     scal.Cast, scal.Switch,
     scal.NEQ, scal.EQ,
     scal.LT, scal.GT, scal.LE, scal.GE,
     scal.Sub, scal.Add, scal.Mod, scal.Mul,
     scal.IntDiv, scal.TrueDiv)
-def get_scalar_constant_value(v):
+def get_scalar_constant_value(orig_v, elemwise=True):
     """return the constant scalar(0-D) value underlying variable `v`
 
     If v is the output of dimshuffles, fills, allocs, rebroadcasts, cast
@@ -523,10 +541,14 @@ def get_scalar_constant_value(v):
     If `v` is not some view of constant scalar data, then raise a
     NotScalarConstantError.
 
+    :param elemwise: If False, we won't try to go into elemwise.
+        So this call is faster.
+
     :note: There may be another function similar to this one in the
         code, but I'm not sure where it is.
     """
-
+    v = orig_v
+    while True:
         if v is None:
             # None is not a scalar (and many uses of this function seem to depend
             # on passing it None)
@@ -535,24 +557,6 @@ def get_scalar_constant_value(v):
         if isinstance(v, (numpy.integer, int, float)):
             return numpy.asarray(v)
 
-    def numpy_scalar(data):
-        """ Return a scalar stored in a numpy ndarray, or raise
-        NotScalarConstantError if the numpy ndarray is not a scalar
-        """
-
-        # handle case where data is numpy.array([])
-        if data.ndim > 0 and  (len(data.shape) == 0 or
-            __builtins__['max'](data.shape) == 0):
-            assert numpy.all(numpy.array([]) == data)
-            raise EmptyConstantError()
-        try:
-            numpy.complex(data)  # works for all numeric scalars
-            return data
-        except Exception:
-            raise NotScalarConstantError(
-                'v.data is non-numeric, non-scalar, or has more than one'
-                ' unique value', data)
-
         if isinstance(v, numpy.ndarray):
             return numpy_scalar(v)
 
@@ -567,9 +571,10 @@ def get_scalar_constant_value(v):
             if isinstance(v.owner.op, (Alloc, DimShuffle, Rebroadcast,
                                        compile.ops.OutputGuard,
                                        compile.DeepCopyOp)):
-            return get_scalar_constant_value(v.owner.inputs[0])
-        elif (isinstance(v.owner.op, theano.compile.ops.Shape_i) and
-            isinstance(v.owner.inputs[0], Constant)):
+                v = v.owner.inputs[0]
+                continue
+            elif isinstance(v.owner.op, theano.compile.ops.Shape_i):
+                if isinstance(v.owner.inputs[0], Constant):
                     return v.owner.inputs[0].data.shape[v.owner.op.i]
             # Don't act as the constant_folding optimization here as this
             # fct is used too early in the optimization phase.  This would
@@ -580,18 +585,20 @@ def get_scalar_constant_value(v):
                 if isinstance(v.owner.op, scal.Second):
                     # We don't need both input to be constant for second
                     shape, val = v.owner.inputs
-                return get_scalar_constant_value(val)
+                    v = val
+                    continue
                 if isinstance(v.owner.op, get_scalar_constant_value_elemwises):
                     const = [get_scalar_constant_value(i)
                              for i in v.owner.inputs]
                     ret = [[None]]
                     v.owner.op.perform(v.owner, const, ret)
                     return ret[0][0]
-        elif isinstance(v.owner.op, Elemwise):
+            elif elemwise and isinstance(v.owner.op, Elemwise):
                 if isinstance(v.owner.op.scalar_op, scal.Second):
                     # We don't need both input to be constant for second
                     shape, val = v.owner.inputs
-                return get_scalar_constant_value(val)
+                    v = val
+                    continue
                 elif isinstance(v.owner.op.scalar_op,
                                 get_scalar_constant_value_elemwises):
                     const = [get_scalar_constant_value(i) for i in v.owner.inputs]
@@ -3075,7 +3082,7 @@ pprint.assign(pow, printing.OperatorPrinter('**', 1, 'right'))
 ##########################
 
 
-def extract_constant(x):
+def extract_constant(x, elemwise=True):
     '''
      This function is basically a call to tensor.get_scalar_constant_value. The
      main difference is the behaviour in case of failure. While
@@ -3085,7 +3092,7 @@ def extract_constant(x):
      ScalarVariable, we convert it to a tensor with tensor_from_scalar.
     '''
     try:
-        x = get_scalar_constant_value(x)
+        x = get_scalar_constant_value(x, elemwise=elemwise)
     except NotScalarConstantError:
         pass
     if (isinstance(x, scal.ScalarVariable) or

theano/tensor/opt.py

@@ -1581,7 +1581,7 @@ def local_upcast_elemwise_constant_inputs(node):
                 else:
                     try:
                         # works only for scalars
-                        cval_i = get_scalar_constant_value(i)
+                        cval_i = get_scalar_constant_value(i, elemwise=False)
                         if all(i.broadcastable):
                             new_inputs.append(T.shape_padleft(
                                 T.cast(cval_i, output_dtype),
@@ -2327,7 +2327,7 @@ def local_remove_switch_const_cond(node):
     """
     if (isinstance(node.op, T.Elemwise) and
         isinstance(node.op.scalar_op, scalar.basic.Switch)):
-        cond = T.extract_constant(node.inputs[0])
+        cond = T.extract_constant(node.inputs[0], elemwise=False)
         if type(cond) is numpy.ndarray and cond.ndim == 0:
             if cond == 0:
                 out = node.inputs[2]
@@ -2377,7 +2377,8 @@ def local_mul_switch_sink(node):
         if i.owner and i.owner.op == T.switch:
             switch = i.owner
             try:
-                if get_scalar_constant_value(switch.inputs[1]) == 0.:
+                if (isinstance(switch.inputs[0], Constant) and
+                    get_scalar_constant_value(switch.inputs[1]) == 0.):
                     listmul = node.inputs[:idx] + node.inputs[idx + 1:]
                     fct = [T.switch(switch.inputs[0], 0,
                                     T.mul(*(listmul + [switch.inputs[2]])))]
@@ -2387,7 +2388,8 @@ def local_mul_switch_sink(node):
             except NotScalarConstantError:
                 pass
             try:
-                if get_scalar_constant_value(switch.inputs[2]) == 0.:
+                if (isinstance(switch.inputs[2], Constant) and
+                    get_scalar_constant_value(switch.inputs[2]) == 0.):
                     listmul = node.inputs[:idx] + node.inputs[idx + 1:]
                     fct = [T.switch(switch.inputs[0],
                                     T.mul(*(listmul + [switch.inputs[1]])), 0)]
@@ -3784,7 +3786,7 @@ def local_abs_merge(node):
         for i in node.inputs:
             if i.owner and i.owner.op == T.abs_:
                 inputs.append(i.owner.inputs[0])
-            else:
+            elif isinstance(i, Constant):
                 try:
                     const = get_scalar_constant_value(i)
                 except NotScalarConstantError:
@@ -3792,6 +3794,8 @@ def local_abs_merge(node):
                 if not (const >= 0).all():
                     return False
                 inputs.append(i)
+            else:
+                return False
         return [T.abs_(T.mul(*inputs))]
     if node.op == T.true_div and sum([i.owner.op == T.abs_ for i in
                                       node.inputs if i.owner]) == 2:

theano/tensor/tests/test_opt.py

@@ -1495,11 +1495,11 @@ def test_log1p():
     f = function([x], T.log(1 + (x)), mode=m)
     assert [node.op for node in f.maker.fgraph.toposort()] == [T.log1p]
     f = function([x], T.log(1 + (-x)), mode=m)
-    assert [node.op for node in f.maker.fgraph.toposort()] == [T.neg,
-         inplace.log1p_inplace]
+    assert [node.op for node in f.maker.fgraph.toposort()] == [
+        T.neg, inplace.log1p_inplace]
     f = function([x], -T.log(1 + (-x)), mode=m)
-    assert [node.op for node in f.maker.fgraph.toposort()] == [T.neg,
-         inplace.log1p_inplace, inplace.neg_inplace]
+    assert [node.op for node in f.maker.fgraph.toposort()] == [
+        T.neg, inplace.log1p_inplace, inplace.neg_inplace]
 
     # check trickier cases (and use different dtype)
     y = fmatrix()
@@ -1507,12 +1507,12 @@ def test_log1p():
     print f.maker.fgraph.toposort()
     # the first three ops are Shape_i, Shape_i, and Dimshuffle
     theano.printing.debugprint(f)
-    assert [node.op for node in f.maker.fgraph.toposort()][3:] \
-            == [T.log1p, tensor.alloc]
+    assert [node.op for node in f.maker.fgraph.toposort()][3:] == [
+        T.log1p, tensor.alloc]
     f = function([x, y], T.log(0 + (x) + tensor.fill(y, 1.0)), mode=m)
     theano.printing.debugprint(f)
-    assert [node.op for node in f.maker.fgraph.toposort()][3:] \
-            == [T.log1p, tensor.alloc]
+    assert [node.op for node in f.maker.fgraph.toposort()][3:] == [
+        T.log1p, tensor.alloc]
     f = function([x, y], T.log(2 + (x) - tensor.fill(y, 1.0)), mode=m)
     theano.printing.debugprint(f)
     assert [node.op for node in f.maker.fgraph.toposort()][3:] \
@@ -2333,7 +2333,8 @@ class Test_alloc_zero(unittest.TestCase):
     def setUp(self):
         mode = theano.compile.mode.get_default_mode()
         self.mode = mode.including("local_incsubtensor_of_allocs",
-             "local_setsubtensor_of_allocs", "local_0_dot_x")
+                                   "local_setsubtensor_of_allocs",
+                                   "local_0_dot_x")
 
     def test_setsubtensor_allocs0(self):
         x = tensor.matrix()
@@ -2427,7 +2428,7 @@ class Test_alloc_zero(unittest.TestCase):
                     f(_e1[1], _e2[1])
                     f(_e1[2], _e2[2])
                     assert numpy.all([not isinstance(x.op, tensor.Dot) for x in
-                                      f.maker.fgraph.toposort() ])
+                                      f.maker.fgraph.toposort()])
 
                     #test that we don't remove shape errors
                     self.assertRaises((ValueError, AssertionError), f,
@@ -2809,8 +2810,8 @@ class test_assert(utt.InferShapeTester):
 
         x = T.scalar()
         y = T.scalar()
-        f = theano.function([x, y], theano.tensor.opt.assert_(x, y,
-             1), mode=mode)
+        f = theano.function([x, y], theano.tensor.opt.assert_(x, y, 1),
+                            mode=mode)
         assert f(1, 1) == 1
         assert f(5, 1) == 5
         topo = f.maker.fgraph.toposort()
@@ -2827,8 +2828,8 @@ class test_assert(utt.InferShapeTester):
 
         x = T.scalar()
         y = T.scalar()
-        f = theano.function([x, y], theano.tensor.opt.assert_(x, y,
-             0), mode=mode)
+        f = theano.function([x, y], theano.tensor.opt.assert_(x, y, 0),
+                            mode=mode)
         self.assertRaises(AssertionError, f, 1, 0)
         topo = f.maker.fgraph.toposort()
         assert len(topo) == 2
@@ -3177,8 +3178,9 @@ def test_constant_get_stabilized():
 class T_local_switch_sink(unittest.TestCase):
     def setUp(self):
         # condition values
-        self.condm = numpy.asarray([[0.1, 0, 1, -1], [0., 0., 0.,
-             0.], [1, 1, 1, 1]])
+        self.condm = numpy.asarray([[0.1, 0, 1, -1],
+                                    [0., 0., 0., 0.],
+                                    [1, 1, 1, 1]])
         self.condv = numpy.asarray([0.1, 0, 1, -1])
         self.conds = [0.1, 0, 1, -1]
 
@@ -3256,14 +3258,14 @@ class T_local_erf(unittest.TestCase):
 
         f = theano.function([x], 1 + T.erf(x), mode=self.mode)
         print f.maker.fgraph.toposort()
-        assert [n.op for n in f.maker.fgraph.toposort()] == [T.mul, T.
-            erfc], f.maker.fgraph.toposort()
+        assert [n.op for n in f.maker.fgraph.toposort()] == [
+            T.mul, T.erfc], f.maker.fgraph.toposort()
         f(val)
 
         f = theano.function([x], T.erf(x) + 1, mode=self.mode)
         print f.maker.fgraph.toposort()
-        assert [n.op for n in f.maker.fgraph.toposort()] == [T.mul, T.
-            erfc], f.maker.fgraph.toposort()
+        assert [n.op for n in f.maker.fgraph.toposort()] == [
+            T.mul, T.erfc], f.maker.fgraph.toposort()
         f(val)
 
         f = theano.function([x], T.erf(x) + 2, mode=self.mode)
@@ -3305,7 +3307,7 @@ class T_local_erf(unittest.TestCase):
         assert topo[0].op == T.erf, f.maker.fgraph.toposort()
         assert isinstance(topo[1].op, T.Elemwise), f.maker.fgraph.toposort()
         assert isinstance(topo[1].op.scalar_op, scal.Add)\
-            or isinstance(topo[1].op.scalar_op,scal.Sub), f.maker.fgraph.toposort()
+            or isinstance(topo[1].op.scalar_op, scal.Sub), f.maker.fgraph.toposort()
         print f(val)
 
     def test_local_erf_minus_one(self):
@@ -3345,7 +3347,8 @@ class T_local_erfc(unittest.TestCase):
             'canonicalize').including('fast_run').excluding('gpu')
         self.mode = self.mode_fusion.excluding('fusion')
         self.mode._optimizer.position_cutoff = 1.50001
-        if theano.config.cxx == '' and not theano.scalar.basic_scipy.imported_scipy_special:
+        if (theano.config.cxx == '' and
+            not theano.scalar.basic_scipy.imported_scipy_special):
             raise SkipTest("erfc need a c++ compiler or scipy")
 
     def test_local_one_minus_erfc(self):