def __call__(self, shape):
if len(shape) != 2:
raise InitializationError("Works only with 2D matrices but shape "
"was: {}".format(shape))
if self.enforce_square and shape[0] != shape[1]:
raise InitializationError("Matrix needs to be square, but was {}"
"".format(shape))
weights = np.eye(shape[0], shape[1], dtype=np.float) * self.scale
weights += self.rnd.randn(*shape) * self.std
return weights
def __call__(self, shape):
if len(shape) != 2:
raise InitializationError("Works only with 2D matrices but shape "
"was: {}".format(shape))
a = self.rnd.randn(*shape)
u, _, v = np.linalg.svd(a, full_matrices=False)
q = u if u.shape == shape else v
return (self.scale * q).reshape(shape)
def __call__(self, shape):
if len(shape) != 2:
raise InitializationError("Works only with 2D matrices but shape "
"was: {}".format(shape))
if shape[0] != shape[1]:
raise InitializationError("Matrix needs to be square, but was {}"
"".format(shape))
N = shape[1]
if self.scale is None:
scale = {
'linear': np.exp(1 / (2 * N)),
'rel': np.sqrt(2) * np.exp(1.2 / (max(N, 6) - 2.4))
}[self.act_func]
else:
scale = self.scale
return scale * self.rnd.randn(*shape) / N
def __call__(self, shape):
self._assert_atleast2d(shape)
if shape[0] != shape[1]:
raise InitializationError("Matrix should be square but was: {}"
"".format(shape))
parameters = self.rnd.uniform(-0.5, 0.5, size=shape)
# normalizing and setting spectral radius (correct, slow):
rho_parameters = max(abs(np.linalg.eig(parameters)[0]))
return parameters * (self.spectral_radius / rho_parameters)
def __call__(self, shape):
self._assert_atleast2d(shape)
if shape[1] < self.connections:
raise InitializationError("Output dimension to small: {} < {}"
"".format(shape[1], self.connections))
sub_result = evaluate_initializer(self.sub_initializer, shape)
connection_mask = np.zeros(shape)
connection_mask[:, :self.connections] = 1.
for i in range(shape[0]):
self.rnd.shuffle(connection_mask[i, :])
return sub_result * connection_mask
def __call__(self, shape):
if shape[0] % 4 != 0:
raise InitializationError("First dim of LstmOpt shape needs to be "
"divisible by 4. But shape was {}"
.format(shape))
weights = np.zeros(shape)
n = shape[0] // 4
sub_shape = (n,) + shape[1:]
weights[:n] = evaluate_initializer(
self.block_input, sub_shape, seed=self.rnd.generate_seed())
weights[n:2 * n] = evaluate_initializer(
self.input_gate, sub_shape, seed=self.rnd.generate_seed())
weights[2 * n:3 * n] = evaluate_initializer(
self.forget_gate, sub_shape, seed=self.rnd.generate_seed())
weights[3 * n:] = evaluate_initializer(
self.output_gate, sub_shape, seed=self.rnd.generate_seed())
return weights
def __call__(self, shape):
if not self.array.shape == shape:
raise InitializationError('Shape mismatch {} != {}'
.format(self.array.shape, shape))
return self.array
def _assert_atleast2d(self, shape):
if len(shape) < 2:
raise InitializationError(
"{} only works on >2D matrices, but shape was {}".format(
self.__class__.__name__, shape))