def grad_check(self):
'''Check whether the full gradient equals to the gradient computed by finite difference at a random point.'''
w = np.random.randn(self.dim)
delta = np.zeros(self.dim)
grad = np.zeros(self.dim)
eps = 1e-4
for i in range(self.dim):
delta[i] = eps
grad[i] = (self.f(w + delta) - self.f(w - delta)) / 2 / eps
delta[i] = 0
if np.linalg.norm(grad - self.grad(w)) > eps:
log.warn('Gradient implementation check failed!')
return False
else:
log.info('Gradient implementation check succeeded!')
return True
def _check_2d_gradient():
res = True
w_2d = xp.random.randn(self.dim, self.n_agent)
g_1d = 0
for i in range(self.n_agent):
g_1d += self.grad(w_2d[:, i], i=i)
g_1d /= self.n_agent
g_2d = self.grad(w_2d).mean(axis=1)
if xp.linalg.norm(g_1d - g_2d) > 1e-5:
log.warn(
'Distributed graident check failed! Difference between global gradient and average of distributed graidents is %.4f'
% xp.linalg.norm(g_1d - g_2d))
res = False
g_2d_sample = self.grad(w_2d,
j=xp.arange(self.m).reshape(-1, 1).repeat(
self.n_agent, axis=1).T).mean(axis=1)
if xp.linalg.norm(g_1d - g_2d_sample) > 1e-5:
log.warn(
'Distributed graident check failed! Difference between global graident and average of all sample gradients is %.4f'
% xp.linalg.norm(g_1d - g_2d_sample))
res = False
samples = xp.random.randint(0, self.m, (self.n_agent, 10))
g_2d_stochastic = self.grad(w_2d, j=samples)
for i in range(self.n_agent):
g_1d_stochastic = self.grad(w_2d[:, i], i=i, j=samples[i])
if xp.linalg.norm(g_1d_stochastic -
g_2d_stochastic[:, i]) > 1e-5:
log.warn(
'Distributed graident check failed! Difference between distributed stoachastic gradient at agent %d and average of sample gradients is %.4f'
% (i,
xp.linalg.norm(g_1d_stochastic -
g_2d_stochastic[:, i])))
res = False
return res
def _check_1d_gradient():
w = xp.random.randn(self.dim)
g = self.grad(w)
g_i = g_ij = 0
res = True
for i in range(self.n_agent):
_tmp_g_i = self.grad(w, i)
_tmp_g_ij = 0
for j in range(self.m):
_tmp_g_ij += self.grad(w, i, j)
if xp.linalg.norm(_tmp_g_i - _tmp_g_ij / self.m) > 1e-5:
log.warn(
'Distributed graident check failed! Difference between local graident at agent %d and average of all local sample gradients is %.4f'
% (i, xp.linalg.norm(_tmp_g_i - _tmp_g_ij / self.m)))
res = False
g_i += _tmp_g_i
g_ij += _tmp_g_ij
g_i /= self.n_agent
g_ij /= self.m_total
if xp.linalg.norm(g - g_i) > 1e-5:
log.warn(
'Distributed gradient check failed! Difference between global graident and average of local gradients is %.4f',
xp.linalg.norm(g - g_i))
res = False
if xp.linalg.norm(g - g_ij) > 1e-5:
log.warn(
'Distributed graident check failed! Difference between global graident and average of all sample gradients is %.4f'
% xp.linalg.norm(g - g_ij))
res = False
return res
def _check_function_value():
w = xp.random.randn(self.dim)
f = self.f(w)
f_i = f_ij = 0
res = True
for i in range(self.n_agent):
_tmp_f_i = self.f(w, i)
_tmp_f_ij = 0
for j in range(self.m):
_tmp_f_ij += self.f(w, i, j)
if xp.abs(_tmp_f_i - _tmp_f_ij / self.m) > 1e-10:
log.warn(
'Distributed function value check failed! Difference between local function value at agent %d and average of all local sample function values %d is %.4f'
% (i, i, xp.abs(_tmp_f_i - _tmp_f_ij / self.m)))
res = False
f_i += _tmp_f_i
f_ij += _tmp_f_ij
f_i /= self.n_agent
f_ij /= self.m_total
if xp.abs(f - f_i) > 1e-10:
log.warn(
'Distributed function value check failed! Difference between the global function value and average of local function values is %.4f'
% xp.abs(f - f_i))
res = False
if xp.abs(f - f_ij) > 1e-10:
log.warn(
'Distributed function value check failed! Difference between the global function value and average of all sample function values is %.4f'
% xp.abs(f - f_ij))
res = False
return res