Merge pull request #5399 from gokul-uf/add_grad_scale

theano/gradient.py

@@ -151,6 +151,8 @@ class DisconnectedType(theano.gof.type.Type):
 
     def __str__(self):
         return 'DisconnectedType'
+
+
 disconnected_type = DisconnectedType()
 
 
@@ -1749,6 +1751,7 @@ Exception args: %s""" % (self.err_pos, self.arg,
                          self.rel_err, self.rel_tol,
                          args_msg)
 
+
 verify_grad.E_grad = GradientError
 
 
@@ -1994,6 +1997,7 @@ class DisconnectedGrad(ViewOp):
     def connection_pattern(self, node):
         return [[False]]
 
+
 disconnected_grad_ = DisconnectedGrad()
 
 
@@ -2062,3 +2066,35 @@ def grad_clip(x, lower_bound, upper_bound):
 
     """
     return GradClip(lower_bound, upper_bound)(x)
+
+
+class GradScale(ViewOp):
+    def __init__(self, multiplier):
+        self.multiplier = multiplier
+
+    def grad(self, args, g_outs):
+        return [self.multiplier * g_out for g_out in g_outs]
+
+
+def grad_scale(x, multiplier):
+    """
+    This op scale or inverse the gradient in the backpropagation.
+
+    :param x: the variable we want its gradient inputs scale
+    :param multiplier: scale of the gradient
+
+    :examples:
+
+        x = theano.tensor.fscalar()
+        fx = theano.tensor.sin(x)
+
+        fp = theano.tensor.grad(fx, wrt=x)
+        fprime = theano.function([x], fp)
+        print(fprime(2))#-0.416
+
+        f_inverse=grad_scale(fx,-1.)
+        fpp = theano.tensor.grad(f_inverse, wrt=x)
+        fpprime = theano.function([x], fpp)
+        print(fpprime(2))#0.416
+    """
+    return GradScale(multiplier)(x)