modif canonicaly to lift toward input abs throught mul and true_div. This allow…

theano/tensor/opt.py

@@ -2010,6 +2010,42 @@ def check_for_x_over_absX(numerators, denominators):
     return numerators, denominators
 local_mul_canonizer.add_simplifier(check_for_x_over_absX, 'X_over_absX')
 
+@register_canonicalize
+@gof.local_optimizer([T.abs_])
+def local_abs_lift(node):
+    """
+    move the abs toward the input. This is needed for check_for_x_over_absX to apply in more case.
+    
+    """
+    if node.op == T.abs_ and node.inputs[0].owner:
+        assert node.nin == 1
+        if node.inputs[0].owner.op == T.mul:
+            return [T.mul(*[T.abs_(i) for i in node.inputs[0].owner.inputs])]
+        if node.inputs[0].owner.op == T.true_div:
+            i = node.inputs[0].owner.inputs
+            return [T.true_div(T.abs_(i[0]),T.abs_(i[1]))]
+    
+@register_specialize
+@gof.local_optimizer([])
+def local_abs_merge(node):
+    """
+    merge abs generated by local_abs_lift when the canonizer don't need it anymore
+    
+    """
+    if node.op == T.mul and sum([i.owner.op == T.abs_ for i in node.inputs if i.owner])>1:
+        inputs = []
+        for i in node.inputs:
+            if i.owner and i.owner.op == T.abs_:
+                inputs.append(i.owner.inputs[0])
+            else:
+                const = get_constant_value(i)
+                if not (const>=0).all():
+                    return False
+                inputs.append(i)
+        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:
+        return [T.abs_(T.true_div(node.inputs[0].owner.inputs[0],node.inputs[1].owner.inputs[0]))]
+
 @register_stabilize
 @gof.local_optimizer([T.log])
 def local_log1p(node):

theano/tensor/tests/test_opt.py

@@ -484,9 +484,54 @@ class test_canonize(unittest.TestCase):
                 assert numpy.all(numpy.isfinite(out))
                 assert numpy.allclose(out,numpy.sign(val_inputs[0]))
                 assert(out_dtype==out.dtype)
+                assert len(f.maker.env.toposort())==1
+
+            #test (2*x) / (3*abs(x)) -> sign(x)
+            for id,(g, sym_inputs, val_inputs, out_dtype) in enumerate([
+                    ((2*dx)/(3*abs(dx)),[dx],[0.5-dxv],'float64'),
+                    ((2*fx)/(3*abs(fx)),[fx],[0.5-fxv],'float32'),
+                    ((2*dx)/(3*abs(dx)),[dx],[0.0*dxv],'float64'),
+                    ((2*fx)/(3*abs(fx)),[fx],[0.0*fxv],'float32'),
+                    ((2*dv)/(3*abs(dv)),[dv],[0.5-dvv],'float64'),
+                    ((2*fv)/(3*abs(fv)),[fv],[0.5-fvv],'float32'),
+                ]):
+                f = compile.function(list(sym_inputs), g,
+                                     mode=mode)
+                topo = f.maker.env.toposort()
+                out = f(*val_inputs)
+                assert numpy.all(numpy.isfinite(out))
+                assert numpy.allclose(out,numpy.sign(val_inputs[0])*2/3)
+                assert(out_dtype==out.dtype)
         finally:
             mode._optimizer = old_optimizer
 
+    def test_abs_mul_div(self):
+        """
+        test that if we have 
+        4 * x / abs(2*x) it get simplifier during canonicalisation.
+        """
+
+        x=T.dscalar()
+        a=T.abs_(x)
+        mode = theano.compile.mode.get_default_mode().excluding("local_elemwise_fusion")
+
+        f=theano.function([x],[(4*x)/abs(2*x)], mode = mode)
+        print f.maker.env.toposort()
+        print
+        f(.1)
+        f(0)
+        f(-1)
+        assert len(f.maker.env.toposort())==2
+
+        f=theano.function([x],[(4*x)/abs(2/x)], mode = mode)
+        print f.maker.env.toposort()
+        print
+        f(.1)
+        f(0)
+        f(-1)
+        assert len(f.maker.env.toposort())==2
+        assert f.maker.env.toposort()[0].op==T.abs_
+
 
     def test_multiple_case_that_fail(self):
         import theano.tensor, theano.compile
@@ -553,6 +598,30 @@ class test_canonize(unittest.TestCase):
         """
         raise SkipTest("Not implemented")
 
+def test_local_merge_abs():
+    x,y,z = T.matrices('xyz')
+    x_val = numpy.random.rand(5,5)
+    y_val = numpy.random.rand(5,5)
+    z_val = numpy.random.rand(5,5)
+    mode = theano.config.mode
+    if mode == "FAST_COMPILE":
+        mode = "FAST_RUN"
+    mode = theano.compile.mode.get_mode(mode).excluding("local_elemwise_fusion")
+
+    f = theano.function([x,y,z],(abs(y*z*-2)), mode=mode)
+    f(x_val,y_val,z_val)
+    theano.printing.debugprint(f)
+    assert isinstance(f.maker.env.toposort()[1].op.scalar_op, scal.Abs)
+    assert len(f.maker.env.toposort())==2
+
+    f = theano.function([x,y,z],abs(x/y), mode=mode)
+    f(x_val,y_val,z_val)
+    theano.printing.debugprint(f)
+    assert isinstance(f.maker.env.toposort()[1].op.scalar_op, scal.Abs)
+    assert len(f.maker.env.toposort())==2
+
+
+
 def test_mixeddiv():
     """Test that int division is preserved"""
     i = iscalar()