def __call__(self, x):
h = self.res_branch2a(x)
h = selu(h)
h = self.res_branch2b(h)
h += x
y = selu(h)
return y
def fwd(self, x):
h_input = F.relu(self.l_input(x))
h0 = F.selu(self.l0(h_input))
h1 = F.selu(self.l1(h0))
h2 = F.selu(self.l2(h1))
h3 = F.selu(self.l3(h2))
h_output = self.l_output(h3)
return h_output
def __call__(self, x):
temp = self.res_branch1(x)
h = self.res_branch2a(x)
h = selu(h)
h = self.res_branch2b(h)
h = temp + h
y = selu(h)
return y
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.selu(x, alpha=self.alpha, scale=self.scale),
x_data,
y_grad,
dtype=numpy.float64,
**self.check_backward_options)
def __call__(self, state):
'''
state: state vector
Q値の範囲が報酬体系によって負の値をとる場合、F.reluは負の値をとれないので、学習に適さない。
活性化関数は、負の値も取ることが可能なものを選択する必要がある。
例えば、F.leaky_relu等。勾配消失問題を考えると、これが良い感じ。
return:
type: Variable of Chainer
Q values of all actions
'''
state32 = state.astype(np.float32)
nrow, ncol = state32.shape
#print('nrow, ncol =', nrow, ncol)
twn_status = chainer.Variable(
state32[:, 0:self.n_size_twn_status].astype(np.float32))
x_ray = state32[:, self.n_size_twn_status:self.n_size_twn_status +
self.num_ray]
eb_status = chainer.Variable(
state32[:, self.n_size_twn_status +
self.num_ray:self.n_size_twn_status + self.num_ray +
self.n_size_eb_status].astype(np.float32))
x = x_ray.reshape(nrow, self.num_ray)
h1 = self.ml1(x)
h3_c = F.concat((twn_status, h1, eb_status), axis=1)
h4 = F.selu(self.l4(h3_c))
h7 = self.ml5(h4)
self.debug_info = (h7, h1, h3_c, h4)
return h7
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.selu(x, alpha=self.alpha, scale=self.scale)
self.assertEqual(y.data.dtype, self.dtype)
expected = numpy.where(self.x >= 0, self.x,
self.alpha * (numpy.exp(self.x) - 1))
expected *= self.scale
testing.assert_allclose(expected, y.data, **self.check_forward_options)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.selu(x, alpha=self.alpha, scale=self.scale)
self.assertEqual(y.data.dtype, self.dtype)
expected = numpy.where(
self.x >= 0, self.x, self.alpha * (numpy.exp(self.x) - 1))
expected *= self.scale
testing.assert_allclose(
expected, y.data, **self.check_forward_options)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.selu(x, alpha=self.alpha, scale=self.scale)
self.assertEqual(y.data.dtype, self.dtype)
expected = self.x.copy()
for i in numpy.ndindex(self.x.shape):
if self.x[i] < 0:
expected[i] = self.alpha * (numpy.exp(expected[i]) - 1)
expected[i] *= self.scale
testing.assert_allclose(expected, y.data, **self.check_forward_options)
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.selu(x, alpha=self.alpha, scale=self.scale)
self.assertEqual(y.data.dtype, self.dtype)
expected = self.x.copy()
for i in numpy.ndindex(self.x.shape):
if self.x[i] < 0:
expected[i] = self.alpha * (numpy.exp(expected[i]) - 1)
expected[i] *= self.scale
testing.assert_allclose(
expected, y.data, **self.check_forward_options)
def check_backward(self, x_data, y_grad):
gradient_check.check_backward(
lambda x: functions.selu(x, alpha=self.alpha, scale=self.scale),
x_data, y_grad,
**self.check_backward_options)
def selu(self, x):
return F.selu(x)
def __call__(self, x):
h = self.conv(x)
y = selu(h)
return y
def __call__(self, x):
h = F.selu(self.l1(x))
h = F.relu(self.l2(h))
return self.l3(h)
def forward(self, inputs, device):
x, = inputs
return functions.selu(x, alpha=self.alpha, scale=self.scale),
def forward(self, inputs, device):
x, = inputs
return functions.selu(x, alpha=self.alpha, scale=self.scale),
