def forward(self, x):
x = self._convert2tensor(x)
self.x = x
self.output = np.tanh(x.value * 0.5) * 0.5 + 0.5
if isinstance(self.x, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self, x, y):
x, y = self._check_input(x, y)
self.x = x
self.y = y
if isinstance(self.x, Constant) and isinstance(self.y, Constant):
return Constant(x.value / y.value)
return Tensor(x.value / y.value, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self._atleast_ndim(x, 2)
self.x = x
if isinstance(self.x, Constant):
return Constant(x.value.flatten())
return Tensor(x.value.flatten(), function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
output = x.value.sum(axis=self.axis, keepdims=self.keepdims)
if isinstance(self.x, Constant):
return Constant(output)
return Tensor(output, function=self)
def forward(self, x, shape):
x = self._convert2tensor(x)
self._atleast_ndim(x, 1)
self.x = x
if isinstance(self.x, Constant):
return Constant(x.value.reshape(*shape))
return Tensor(x.value.reshape(*shape), function=self)
def forward(self, x, shape):
x = self._convert2tensor(x)
self.x = x
output = np.broadcast_to(x.value, shape)
if isinstance(self.x, Constant):
return Constant(output)
return Tensor(output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
self.output = sp.gamma(x.value)
if isinstance(x, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self):
eps = 0.5 - np.random.uniform(size=self.loc.shape)
self.eps = np.sign(eps) * np.log(1 - 2 * np.abs(eps))
self.output = self.loc.value - self.scale.value * self.eps
if isinstance(self.loc, Constant) and isinstance(self.scale, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
output = x.value.clip(min=0)
if isinstance(x, Constant):
return Constant(output)
return Tensor(output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
output = np.log(self.x.value)
if isinstance(self.x, Constant):
return Constant(output)
return Tensor(output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
output = np.maximum(x.value, 0) + np.log1p(np.exp(-np.abs(x.value)))
if isinstance(x, Constant):
return Constant(output)
return Tensor(output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
self.output = np.linalg.cholesky(x.value)
if isinstance(self.x, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self, x, mu, tau):
x, mu, tau = self._check_input(x, mu, tau)
self.x = x
self.mu = mu
self.tau = tau
output = (-0.5 * np.square(x.value - mu.value) * tau.value +
0.5 * np.log(tau.value) - 0.5 * np.log(2 * np.pi))
return Tensor(output, function=self)
def forward(self, x, y):
x, y = self._check_input(x, y)
self.x = x
self.y = y
self.output = np.power(x.value, y.value)
if isinstance(self.x, Constant) and isinstance(self.y, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def _forward(self, x, istraining=False):
x = self._convert2tensor(x)
if istraining:
self.x = x
self.mask = (np.random.rand(*x.shape) > self.prob) * self.coef
return Tensor(x.value * self.mask, function=self)
else:
return x
def forward(self, x):
x = self._convert2tensor(x)
if self.axes is not None:
self._equal_ndim(x, len(self.axes))
self.x = x
if isinstance(self.x, Constant):
return Constant(np.transpose(x.value, self.axes))
return Tensor(np.transpose(x.value, self.axes), function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
self.output = np.abs(x.value)
if isinstance(x, Constant):
return Constant(self.output)
self.sign = np.sign(x.value)
return Tensor(self.output, function=self)
def forward(self, x, t):
x, t = self._check_input(x, t)
self.x = x
self.t = t
self.y = self._softmax(x.value)
np.clip(self.y, 1e-10, 1, out=self.y)
loss = -t.value * np.log(self.y)
return Tensor(loss, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
self._equal_ndim(x, 2)
self.output = np.linalg.inv(x.value)
if isinstance(self.x, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self, x, t):
x, t = self._check_input(x, t)
self.x = x
self.t = t
# y = self.forward(x)
# np.clip(y, 1e-10, 1 - 1e-10, out=y)
# return np.sum(-t * np.log(y) - (1 - t) * np.log(1 - y))
loss = (np.maximum(x.value, 0) - t.value * x.value +
np.log1p(np.exp(-np.abs(x.value))))
return Tensor(loss, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
self._equal_ndim(x, 2)
sign, self.output = np.linalg.slogdet(x.value)
if sign != 1:
raise ValueError("matrix has to be positive-definite")
if isinstance(self.x, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self._atleast_ndim(x, 1)
self.x = x
output = np.split(x.value, self.indices_or_sections, self.axis)
if isinstance(self.x, Constant):
return tuple([Constant(out) for out in output])
self.n_output = len(output)
self.delta = [None for _ in output]
return tuple([Tensor(out, function=self) for out in output])
def forward(self, x, y):
x, y = self._check_input(x, y)
self.x = x
self.y = y
img = np.pad(x.value, [(p, ) for p in self.pad], "constant")
self.shape = img.shape
self.patch = img2patch(img, y.shape[:2], self.stride)
return Tensor(np.tensordot(self.patch,
y.value,
axes=((3, 4, 5), (0, 1, 2))),
function=self)
def forward(self):
if self.coef.ndim != 1:
raise NotImplementedError
indices = np.array(
[np.random.choice(self.n_component, p=c) for c in self.coef.value])
output = np.random.normal(loc=self.mu.value[indices],
scale=self.std.value[indices])
if (isinstance(self.coef, Constant) and isinstance(self.mu, Constant)
and isinstance(self.std, Constant)):
return Constant(output)
return Tensor(output, function=self)
def forward(self, a, b):
a = self._convert2tensor(a)
b = self._convert2tensor(b)
self._equal_ndim(a, 2)
self._equal_ndim(b, 2)
self.a = a
self.b = b
self.output = np.linalg.solve(a.value, b.value)
if isinstance(self.a, Constant) and isinstance(self.b, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self._equal_ndim(x, 4)
self.x = x
img = np.pad(x.value, [(p, ) for p in self.pad], "constant")
patch = img2patch(img, self.pool_size, self.stride)
n_batch, xlen_out, ylen_out, _, _, in_channels = patch.shape
patch = patch.reshape(n_batch, xlen_out, ylen_out, -1, in_channels)
self.shape = img.shape
self.index = patch.argmax(axis=3)
return Tensor(patch.max(axis=3), function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
self.output = np.prod(self.x.value, axis=self.axis, keepdims=True)
if not self.keepdims:
output = np.squeeze(self.output)
if output.size == 1:
output = output.item()
else:
output = self.output
if isinstance(self.x, Constant):
return Constant(output)
return Tensor(output, function=self)
def forward(self):
self.eps = np.random.standard_cauchy(size=self.loc.shape)
self.output = self.scale.value * self.eps + self.loc.value
if isinstance(self.loc, Constant):
return Constant(self.output)
return Tensor(self.output, function=self)
def forward(self, x):
x = self._convert2tensor(x)
self.x = x
if isinstance(self.x, Constant):
return Constant(np.square(x.value))
return Tensor(np.square(x.value), function=self)
def _forward(self, x):
x = self._convert2tensor(x)
self.x = x
self.output = x.value - self._logsumexp(x.value)
return Tensor(self.output, function=self)