def test_softmax2(self):
x = np.ones(10)
x[0] = 2
self.assertTrue(np.argmax(softmax(x)) == 0)
x[0] = 1
x[5] = 2
self.assertTrue(np.argmax(softmax(x)) == 5)
def get_matrix_and_derivatives(self, x):
"""Produces the circuit matrix and partials for this gate."""
alpha, l = self.partition_input(x)
l = utils.softmax(l, 10)
Hv = []
stride = self.working_sigmav.shape[1]
for i in range(self.working_sigmav.shape[0]):
Hv.append(
utils.dot_product(alpha[i * stride:(i + 1) * stride],
self.working_sigmav[i]))
H = utils.dot_product(l, Hv)
# Partials of H with respect to function variables
alpha_der = np.array([
l[i] * self.working_sigmav[i]
for i in range(self.get_location_count())
])
alpha_der = alpha_der.reshape(
(-1, alpha_der.shape[-2], alpha_der.shape[-1]))
# Partials of H with respect to location variables
L = np.tile(l, (self.working_sigmav.shape[0], 1))
L = np.identity(self.working_sigmav.shape[0]) - L
L = 10 * (np.diag(l) @ L)
l_der = np.array([utils.dot_product(Lr, Hv) for Lr in L])
_, dav = utils.dexpmv(H, alpha_der)
_, dlv = utils.dexpmv(H, l_der)
return sp.linalg.expm(H), np.concatenate([dav, dlv])
def get_matrix ( self, x ):
"""Produces the circuit matrix for this gate."""
alpha, l = self.partition_input( x )
l = utils.softmax( l, 10 )
H = utils.dot_product( alpha, self.sigmav )
U = sp.linalg.expm( H )
P = utils.dot_product( l, self.working_perms )
return P @ np.kron( U, self.I ) @ P.T
def get_matrix(self, x):
"""Produces the circuit matrix for this gate."""
alpha, l = self.partition_input(x)
l = utils.softmax(l, 10)
Hv = []
stride = self.working_sigmav.shape[1]
for i in range(self.working_sigmav.shape[0]):
Hv.append(
utils.dot_product(alpha[i * stride:(i + 1) * stride],
self.working_sigmav[i]))
H = utils.dot_product(l, Hv)
return sp.linalg.expm(H)
def get_matrix_and_derivatives ( self, x ):
"""Produces the circuit matrix and partials for this gate."""
alpha, l = self.partition_input( x )
l = utils.softmax( l, 10 )
H = utils.dot_product( alpha, self.sigmav )
P = utils.dot_product( l, self.working_perms )
U = np.kron( sp.linalg.expm( H ), self.I )
PU = P @ U
UP = U @ P.T
PUP = PU @ P.T
_, dav = utils.dexpmv( H, self.sigmav )
dav = np.kron( dav, self.I )
dav = P @ dav @ P.T
dlv = self.working_perms @ UP + PU @ self.working_perms.transpose( ( 0, 2, 1 ) ) - 2*PUP
dlv = np.array( [ x*y for x, y in zip( 10*l, dlv ) ] )
return PUP, np.concatenate( [ dav, dlv ] )
def test_softmax1(self):
for i in range(10):
x = 10 * np.random.random(100)
self.assertTrue(np.abs(np.sum(softmax(x)) - 1) < 1e-15)