def test_top_sparse3(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
model = utils.parafac(factors)
for beta in [1.0, 2.0, 1.5]:
for factor in range(3):
# Generate the top numerator for the reference dense method
einstr = utils.generate_dense(rank, factor)
# Get all factors that aren't the current factor
mode_factors = [a for j, a in enumerate(factors) if j != factor]
mode_factors += [x * (model ** (beta - 2.)), ]
top_d = np.einsum(einstr, *mode_factors)
# Now get the top numerator for the sparse method
top_s = np.zeros(factors[factor].shape, dtype=np.float32)
utils.top_sparse3(x_indices, x_vals, top_s, beta, factor, *factors)
# Now get the top numerator for the sparse method using numba
top_n = np.zeros(factors[factor].shape, dtype=np.float32)
utils.top_sparse3_numba(x_indices, x_vals, top_n, beta, factor, *factors)
print(top_d, top_s, top_n)
print(top_d - top_n)
print(np.histogram(np.abs((top_d - top_n) / top_d)))
result = np.allclose(top_d, top_s, rtol=1e-2, atol=1e-2)
self.assertTrue(result)
result = np.allclose(top_d, top_n, rtol=1e-2, atol=1e-2)
self.assertTrue(result)
def test_bot_sparse3(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
model = utils.parafac(factors)
for beta in [1.0, 2.0, 1.5]:
for factor in range(3):
# Generate the bottom denominator for the reference dense method
einstr = utils.generate_dense(rank, factor)
# Get all factors that aren't the current factor
mode_factors = [a for j, a in enumerate(factors) if j != factor]
mode_factors += [(model ** (beta - 1.)), ]
bot_d = np.einsum(einstr, *mode_factors)
# Now get the bottom denominator for the sparse method
bot_s = np.zeros(factors[factor].shape, dtype=np.float32)
utils.bot_sparse3(x_indices, x_vals, bot_s, beta, factor, *factors)
# Now get the bottom denominator for the sparse method
bot_n = np.zeros(factors[factor].shape, dtype=np.float32)
utils.bot_sparse3(x_indices, x_vals, bot_n, beta, factor, *factors)
result = np.allclose(bot_d, bot_s, rtol=1e-2)
self.assertTrue(result)
result = np.allclose(bot_d, bot_n, rtol=1e-2)
self.assertTrue(result)
def test_top_sparse3(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
model = utils.parafac(factors)
for beta in [1.0, 2.0, 1.5]:
for factor in range(3):
# Generate the top numerator for the reference dense method
einstr = utils.generate_dense(rank, factor)
# Get all factors that aren't the current factor
mode_factors = [
a for j, a in enumerate(factors) if j != factor
]
mode_factors += [
x * (model**(beta - 2.)),
]
top_d = np.einsum(einstr, *mode_factors)
# Now get the top numerator for the sparse method
top_s = np.zeros(factors[factor].shape, dtype=np.float32)
utils.top_sparse3(x_indices, x_vals, top_s, beta, factor,
*factors)
# Now get the top numerator for the sparse method using numba
top_n = np.zeros(factors[factor].shape, dtype=np.float32)
utils.top_sparse3_numba(x_indices, x_vals, top_n, beta, factor,
*factors)
print(top_d, top_s, top_n)
print(top_d - top_n)
print(np.histogram(np.abs((top_d - top_n) / top_d)))
result = np.allclose(top_d, top_s, rtol=1e-2, atol=1e-2)
self.assertTrue(result)
result = np.allclose(top_d, top_n, rtol=1e-2, atol=1e-2)
self.assertTrue(result)
def test_bot_sparse3(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
model = utils.parafac(factors)
for beta in [1.0, 2.0, 1.5]:
for factor in range(3):
# Generate the bottom denominator for the reference dense method
einstr = utils.generate_dense(rank, factor)
# Get all factors that aren't the current factor
mode_factors = [
a for j, a in enumerate(factors) if j != factor
]
mode_factors += [
(model**(beta - 1.)),
]
bot_d = np.einsum(einstr, *mode_factors)
# Now get the bottom denominator for the sparse method
bot_s = np.zeros(factors[factor].shape, dtype=np.float32)
utils.bot_sparse3(x_indices, x_vals, bot_s, beta, factor,
*factors)
# Now get the bottom denominator for the sparse method
bot_n = np.zeros(factors[factor].shape, dtype=np.float32)
utils.bot_sparse3(x_indices, x_vals, bot_n, beta, factor,
*factors)
result = np.allclose(bot_d, bot_s, rtol=1e-2)
self.assertTrue(result)
result = np.allclose(bot_d, bot_n, rtol=1e-2)
self.assertTrue(result)
def test_beta_divergence(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
for beta in [1, 1.5, 2]:
a = x
b = utils.parafac(factors)
div_d = utils.beta_divergence_dense(a, b, beta)
div_s = utils.beta_divergence(x_indices, x_vals, beta, *factors)
div_s.shape = div_d.shape
close = np.allclose(div_s, div_d, rtol=1e-5)
self.assertTrue(close)
def test_beta_divergence(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
for beta in [1, 1.5, 2]:
a = x
b = utils.parafac(factors)
div_d = utils.beta_divergence_dense(a, b, beta)
div_s = utils.beta_divergence(x_indices, x_vals, beta, *factors)
div_s.shape = div_d.shape
close = np.allclose(div_s, div_d, rtol=1e-5)
self.assertTrue(close)
def test_fit(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
bnf = betantf.BetaNTF(shape, n_components=k)
before = bnf.score(x_indices, x_vals)
initial = bnf.impute(x_indices)
initial.shape = x.shape
reconstructed = bnf.fit(x_indices, x_vals)
reconstructed.shape = x.shape
after = bnf.score()
self.assertTrue(after < before)
e_0 = rmse(x, initial)
e_1 = rmse(x, reconstructed)
self.assertTrue(e_0 > e_1)
def test_fit(self):
shape, rank, k, factors, x, x_indices, x_vals = \
utils.generate_dataset()
bnf = betantf.BetaNTF(shape, n_components=k)
before = bnf.score(x_indices, x_vals)
initial = bnf.impute(x_indices)
initial.shape = x.shape
reconstructed = bnf.fit(x_indices, x_vals)
reconstructed.shape = x.shape
after = bnf.score()
self.assertTrue(after < before)
e_0 = rmse(x, initial)
e_1 = rmse(x, reconstructed)
self.assertTrue(e_0 > e_1)
