def _value_and_gradient(fn, *args):
"""Calls `fn` and computes the gradient of the result wrt `arg`."""
if tfe_context.executing_eagerly():
v, g = tfe_backprop.val_and_grad_function(fn)(args)
else:
v = fn(*args)
g = gradients_impl.gradients(v, args)
return v, g
def _value_and_gradient(fn, *args):
"""Calls `fn` and computes the gradient of the result wrt `arg`."""
if tfe_context.executing_eagerly():
v, g = tfe_backprop.val_and_grad_function(fn)(args)
else:
v = fn(*args)
g = gradients_impl.gradients(v, args)
return v, g
def testNonEmptyParamsForValueAndGradFunction(self):
def fn(a, b):
return a * b
val_and_grad_fn = backprop.val_and_grad_function(fn, params=[1])
x = 2.0
y = 3.0
val, grads = val_and_grad_fn(x, y)
self.assertAllClose(val, x * y)
self.assertEqual(1, len(grads))
self.assertAllEqual(grads[0], x)
def testEmptyParamsForValueAndGradFunction(self):
def fn(a, b):
return a * b
val_and_grads_fn = backprop.val_and_grad_function(fn)
x = 2.0
y = 3.0
val, (dx, dy) = val_and_grads_fn(x, y)
self.assertAllClose(val, x * y)
self.assertAllEqual(dx, y)
self.assertAllEqual(dy, x)
def testNonEmptyParamsForValueAndGradFunction(self):
def fn(a, b):
return a * b
val_and_grad_fn = backprop.val_and_grad_function(fn, params=[1])
x = 2.0
y = 3.0
val, grads = val_and_grad_fn(x, y)
self.assertAllClose(val, x * y)
self.assertEqual(1, len(grads))
self.assertAllEqual(grads[0], x)
def testEmptyParamsForValueAndGradFunction(self):
def fn(a, b):
return a * b
val_and_grads_fn = backprop.val_and_grad_function(fn)
x = 2.0
y = 3.0
val, (dx, dy) = val_and_grads_fn(x, y)
self.assertAllClose(val, x * y)
self.assertAllEqual(dx, y)
self.assertAllEqual(dy, x)
def testBasicFunctionalWithValue(self):
def forward(a, b):
mm = math_ops.matmul(a, b)
return math_ops.reduce_sum(mm)
aa = constant_op.constant([[1., 0.], [0., 1.]])
bb = constant_op.constant([[1., 2.], [3., 4.]])
val, (da, ) = backprop.val_and_grad_function(forward, ['a'])(aa, bb)
self.assertAllEqual(
da,
math_ops.matmul(array_ops.ones_like(aa), array_ops.transpose(bb)))
self.assertAllEqual(val, forward(aa, bb))
def testBasicFunctionalWithValue(self):
def forward(a, b):
mm = math_ops.matmul(a, b)
return math_ops.reduce_sum(mm)
aa = constant_op.constant([[1., 0.], [0., 1.]])
bb = constant_op.constant([[1., 2.], [3., 4.]])
val, (da,) = backprop.val_and_grad_function(forward, ['a'])(aa, bb)
self.assertAllEqual(da,
math_ops.matmul(
array_ops.ones_like(aa),
array_ops.transpose(bb)))
self.assertAllEqual(val, forward(aa, bb))
def testDifferentiatingFunctionThatReturnsNone(self):
def fn(x, y):
result = x*y # pylint: disable=unused-variable
x = constant_op.constant(1)
y = constant_op.constant(2)
loss_grads_fn = backprop.implicit_val_and_grad(fn)
with self.assertRaisesRegexp(
ValueError, 'Cannot differentiate a function that returns None; '
'did you forget to return a value from fn?'):
loss_grads_fn(x, y)
val_and_grads_fn = backprop.val_and_grad_function(fn)
with self.assertRaisesRegexp(
ValueError, 'Cannot differentiate a function that returns None; '
'did you forget to return a value from fn?'):
val_and_grads_fn(x, y)
def testDifferentiatingFunctionThatReturnsNone(self):
def fn(x, y):
result = x*y # pylint: disable=unused-variable
x = constant_op.constant(1)
y = constant_op.constant(2)
loss_grads_fn = backprop.implicit_val_and_grad(fn)
with self.assertRaisesRegexp(
ValueError, 'Cannot differentiate a function that returns None; '
'did you forget to return a value from fn?'):
loss_grads_fn(x, y)
val_and_grads_fn = backprop.val_and_grad_function(fn)
with self.assertRaisesRegexp(
ValueError, 'Cannot differentiate a function that returns None; '
'did you forget to return a value from fn?'):
val_and_grads_fn(x, y)
