def test_admm_with_large_lamduh(N, p, nchunks):
X = da.random.random((N, p), chunks=(N // nchunks, p))
beta = np.random.random(p)
y = make_y(X, beta=np.array(beta), chunks=(N // nchunks, ))
X, y = persist(X, y)
z = admm(X, y, reg=L1, lamduh=1e4, rho=20, max_iter=500)
assert np.allclose(z, np.zeros(p), atol=1e-4)
def test_admm_with_large_lamduh(N, p, nchunks, is_cupy):
X = da.random.random((N, p), chunks=(N // nchunks, p))
beta = np.random.random(p)
y = make_y(X, beta=np.array(beta), chunks=(N // nchunks, ))
if is_cupy:
cupy = pytest.importorskip('cupy')
X, y = to_dask_cupy_array_xy(X, y, cupy)
X, y = persist(X, y)
z = admm(X, y, regularizer=L1(), lamduh=1e5, rho=20, max_iter=500)
assert np.allclose(z, np.zeros(p), atol=1e-4)
def test_basic_unreg_descent(func, kwargs, N, nchunks, family):
beta = np.random.normal(size=2)
M = len(beta)
X = da.random.random((N, M), chunks=(N // nchunks, M))
y = make_y(X, beta=np.array(beta), chunks=(N // nchunks, ))
X, y = persist(X, y)
result = func(X, y, family=family, **kwargs)
test_vec = np.random.normal(size=2)
opt = family.pointwise_loss(result, X, y).compute()
test_val = family.pointwise_loss(test_vec, X, y).compute()
assert opt < test_val
def make_intercept_data(N, p, seed=20009):
'''Given the desired number of observations (N) and
the desired number of variables (p), creates
random logistic data to test on.'''
# set the seeds
da.random.seed(seed)
np.random.seed(seed)
X = np.random.random((N, p + 1))
col_sums = X.sum(axis=0)
X = X / col_sums[None, :]
X[:, p] = 1
X = da.from_array(X, chunks=(N / 5, p + 1))
y = make_y(X, beta=np.random.random(p + 1))
return X, y
def test_basic_reg_descent(func, kwargs, N, nchunks, family, lam, reg):
beta = np.random.normal(size=2)
M = len(beta)
X = da.random.random((N, M), chunks=(N // nchunks, M))
y = make_y(X, beta=np.array(beta), chunks=(N // nchunks, ))
X, y = persist(X, y)
result = func(X, y, family=family, lamduh=lam, regularizer=reg, **kwargs)
test_vec = np.random.normal(size=2)
f = reg.add_reg_f(family.pointwise_loss, lam)
opt = f(result, X, y).compute()
test_val = f(test_vec, X, y).compute()
assert opt < test_val
def test_basic_unreg_descent(func, kwargs, N, nchunks, family, is_cupy):
beta = np.random.normal(size=2)
M = len(beta)
X = da.random.random((N, M), chunks=(N // nchunks, M))
y = make_y(X, beta=np.array(beta), chunks=(N // nchunks, ))
if is_cupy:
cupy = pytest.importorskip('cupy')
X, y = to_dask_cupy_array_xy(X, y, cupy)
X, y = persist(X, y)
result = func(X, y, family=family, **kwargs)
test_vec = np.random.normal(size=2)
test_vec = maybe_to_cupy(test_vec, X)
opt = family.pointwise_loss(result, X, y).compute()
test_val = family.pointwise_loss(test_vec, X, y).compute()
assert opt < test_val
N = 100000
S = 2
X = np.random.randn(N, M)
X[:, 1] = 1.0
beta0 = np.random.randn(M)
def make_y(X, beta0=beta0):
N, M = X.shape
z0 = X.dot(beta0)
z0 = da.compute(z0)[0] # ensure z0 is a numpy array
scl = S / z0.std()
beta0 *= scl
z0 *= scl
y = np.random.rand(N) < logit(z0)
return y, z0
y, z0 = make_y(X)
L0 = N * math.log(2.0)
dX = da.from_array(X, chunks=(N / 10, M))
dy = da.from_array(y, chunks=(N / 10,))
@pytest.mark.parametrize('X,y', [(dX, dy)])
def test_gradient(X, y):
gradient(X, y)
