def _run_minibatch(self, y, mask):
D = self.D.copy()
start_iter = 31
it, D, x = nmf.solve(
self.y, D, x=None, tol=1.0e-6,
minibatch=30, maxiter=start_iter,
method=self.method,
likelihood=self.likelihood, mask=mask,
random_seed=0)
start_error = self.error(x, D, mask=mask)
assert not allclose(x, xp.zeros_like(x), atol=1.0e-5)
# make sure the iteration reduces the error
errors = []
for i in range(10):
print('trial ', i)
it, D, x = nmf.solve(
self.y, D, x=x, tol=1.0e-6,
minibatch=30, maxiter=self.maxiter,
method=self.method,
likelihood=self.likelihood, mask=mask,
random_seed=0)
# just make sure the loss is decreasing
error = self.error(x, D, mask=mask)
assert not allclose(x, xp.zeros_like(x), atol=1.0e-5)
assert error < start_error
errors.append(error)
# make sure errors are decreasing in average
errors = np.array(errors)
assert np.mean(errors[:4]) > np.mean(errors[-4:])
def test_minibatch_mask(self, method, all_xp=False):
D = xp.array(self.D)
y = xp.array(self.y) if all_xp else self.y
mask = xp.array(self.mask) if all_xp else self.mask
it, D, x = nmf.solve(y, D, tol=1.0e-10,
method=method, maxiter=self.maxiter,
mask=mask, minibatch=self.minibatch)
def run_fullbatch(self, mask):
D = self.D.copy()
it, D, x = nmf.solve(self.y, D, x=None, tol=1.0e-6,
minibatch=None, maxiter=self.maxiter,
method=self.method,
likelihood=self.likelihood, mask=mask,
random_seed=0)
assert it < self.maxiter - 1
self.assert_minimum(x, D, tol=1.0e-5, n=100, mask=mask)
assert not allclose(x, xp.zeros_like(x), atol=1.0e-5)
def test_mask(self, method):
it, D, x = nmf.solve(xp.array(self.y), xp.array(self.D), tol=1.0e-10,
method=method, maxiter=self.maxiter,
mask=xp.array(self.mask))
def test_minibatch(self, method, all_xp=False):
D = xp.array(self.D, dtype=np.float32)
y = xp.array(self.y, dtype=np.float32) if all_xp else self.y
it, D, x = nmf.solve(y, D, tol=1.0e-10,
method=method, maxiter=self.maxiter,
minibatch=self.minibatch)