def test_not_keep_graph(self):
x = chainer.Variable(numpy.array([1], numpy.float32))
y = functions.Sigmoid()(x)
reporter = chainer.Reporter()
with self._scope(False):
reporter.report({'y': y})
self.assertIsNone(reporter.observation['y'].creator)
def test_keep_graph_default(self):
x = chainer.Variable(numpy.array([1], numpy.float32))
y, = functions.Sigmoid().apply((x, ))
reporter = chainer.Reporter()
with self._scope(None):
reporter.report({'y': y})
self.assertIsNone(reporter.observation['y'].creator)
def backward(self, inputs, grad_outputs):
xp = cuda.get_array_module(*inputs)
x, t = inputs
gloss = grad_outputs[0]
y, = f.Sigmoid(self.use_cudnn).forward((x, ))
gx = gloss * (y - t)
return gx, None
def __call__(self, x):
self.flag_history.append(configuration.config.keep_graph_on_report)
h = self.l1(x)
if self.i == 0:
h, = functions.Sigmoid().apply((h, ))
else:
h = functions.Tanh()(h)
h = self.l2(h)
self.i += 1
return h
def func(self, c, h):
# self.a1 = numpy.tanh(a1)
# self.i1 = _sigmoid(i1)
# self.f1 = _sigmoid(f1)
#
# self.a2 = numpy.tanh(a2)
# self.i2 = _sigmoid(i2)
# self.f2 = _sigmoid(f2)
#
# self.o = _sigmoid(o1 + o2)
# self.c = self.a1 * self.i1 + self.a2 * self.i2 + \
# self.f1 * c_prev1 + self.f2 * c_prev2
#
# h = self.o * numpy.tanh(self.c)
a1, i1, f1, o1 = chainer_extract_gates(h)
a2, i2, f2, o2 = chainer_extract_gates(h)
self.c = F.Tanh(a1) * F.Sigmoid(i1) + F.Sigmoid(f1) * c_prev1 + \
F.Tanh(a2) * F.Sigmoid(i2) + F.Sigmoid(f2) * c_prev2
self.o = F.Sigmoid(o1 + o2)
h = self.o * numpy.tanh(self.c)
def check_backward(self, x_data, y_grad, use_cudnn=True):
gradient_check.check_backward(functions.Sigmoid(use_cudnn), x_data,
y_grad)
def check_backward(self, x_data, y_grad, use_cudnn='always'):
with chainer.using_config('use_cudnn', use_cudnn):
gradient_check.check_backward(functions.Sigmoid(), x_data, y_grad,
**self.check_backward_options)
