def test_conditional_loglikelihood(self):
mcgsm = MCGSM(3, 1, 2, 1, 4)
mcgsm.linear_features = randn(mcgsm.num_components, mcgsm.dim_in) / 5.
mcgsm.means = randn(mcgsm.dim_out, mcgsm.num_components) / 5.
M = 100
inputs = randn(mcgsm.dim_in, M)
outputs = mcgsm.sample(inputs)
loglik0 = mcgsm.loglikelihood(inputs, outputs)
loglik1 = []
N = 1000
# estimate log-likelihood via sampling
for _ in range(N):
labels = mcgsm.sample_prior(inputs)
loglik1.append(mcgsm.loglikelihood(inputs, outputs, labels))
loglik1 = vstack(loglik1)
d = abs(logmeanexp(loglik1, 0) - loglik0).ravel()
s = std(loglik1, 0, ddof=1).ravel()
for i in range(M):
self.assertLess(d[i], 6. * s[i] / sqrt(N))
def test_pickle(self):
mcgsm0 = MCGSM(11, 2, 4, 7, 21)
mcgsm0.linear_features = randn(mcgsm0.num_components, mcgsm0.dim_in)
mcgsm0.means = randn(mcgsm0.dim_out, mcgsm0.num_components)
tmp_file = mkstemp()[1]
# store model
with open(tmp_file, "w") as handle:
dump({"mcgsm": mcgsm0}, handle)
# load model
with open(tmp_file) as handle:
mcgsm1 = load(handle)["mcgsm"]
# make sure parameters haven't changed
self.assertEqual(mcgsm0.dim_in, mcgsm1.dim_in)
self.assertEqual(mcgsm0.dim_out, mcgsm1.dim_out)
self.assertEqual(mcgsm0.num_components, mcgsm1.num_components)
self.assertEqual(mcgsm0.num_scales, mcgsm1.num_scales)
self.assertEqual(mcgsm0.num_features, mcgsm1.num_features)
self.assertLess(max(abs(mcgsm0.scales - mcgsm1.scales)), 1e-20)
self.assertLess(max(abs(mcgsm0.weights - mcgsm1.weights)), 1e-20)
self.assertLess(max(abs(mcgsm0.features - mcgsm1.features)), 1e-20)
self.assertLess(max(abs(mcgsm0.linear_features - mcgsm1.linear_features)), 1e-20)
self.assertLess(max(abs(mcgsm0.means - mcgsm1.means)), 1e-20)
for chol0, chol1 in zip(mcgsm0.cholesky_factors, mcgsm1.cholesky_factors):
self.assertLess(max(abs(chol0 - chol1)), 1e-20)
for pred0, pred1 in zip(mcgsm0.predictors, mcgsm1.predictors):
self.assertLess(max(abs(pred0 - pred1)), 1e-20)
def test_conditional_loglikelihood(self):
mcgsm = MCGSM(3, 1, 2, 1, 4)
mcgsm.linear_features = randn(mcgsm.num_components, mcgsm.dim_in) / 5.0
mcgsm.means = randn(mcgsm.dim_out, mcgsm.num_components) / 5.0
M = 100
inputs = randn(mcgsm.dim_in, M)
outputs = mcgsm.sample(inputs)
loglik0 = mcgsm.loglikelihood(inputs, outputs)
loglik1 = []
N = 1000
# estimate log-likelihood via sampling
for _ in range(N):
labels = mcgsm.sample_prior(inputs)
loglik1.append(mcgsm.loglikelihood(inputs, outputs, labels))
loglik1 = vstack(loglik1)
d = abs(logmeanexp(loglik1, 0) - loglik0).ravel()
s = std(loglik1, 0, ddof=1).ravel()
for i in range(M):
self.assertLess(d[i], 6.0 * s[i] / sqrt(N))
def test_gradient(self):
mcgsm = MCGSM(5, 2, 2, 4, 10)
cholesky_factors = []
for k in range(mcgsm.num_components):
cholesky_factors.append(cholesky(cov(randn(mcgsm.dim_out, mcgsm.dim_out ** 2))))
mcgsm.cholesky_factors = cholesky_factors
mcgsm.linear_features = randn(mcgsm.num_components, mcgsm.dim_in) / 5.0
mcgsm.means = randn(mcgsm.dim_out, mcgsm.num_components) / 5.0
err = mcgsm._check_gradient(randn(mcgsm.dim_in, 1000), randn(mcgsm.dim_out, 1000), 1e-5)
self.assertLess(err, 1e-8)
# without regularization
for param in ["priors", "scales", "weights", "features", "chol", "pred", "linear_features", "means"]:
err = mcgsm._check_gradient(
randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000),
1e-5,
parameters={
"train_prior": param == "priors",
"train_scales": param == "scales",
"train_weights": param == "weights",
"train_features": param == "features",
"train_cholesky_factors": param == "chol",
"train_predictors": param == "pred",
"train_linear_features": param == "linear_features",
"train_means": param == "means",
},
)
self.assertLess(err, 1e-8)
# with regularization
for norm in ["L1", "L2"]:
for param in ["priors", "scales", "weights", "features", "chol", "pred", "linear_features", "means"]:
err = mcgsm._check_gradient(
randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000),
1e-7,
parameters={
"train_prior": param == "priors",
"train_scales": param == "scales",
"train_weights": param == "weights",
"train_features": param == "features",
"train_cholesky_factors": param == "chol",
"train_predictors": param == "pred",
"train_linear_features": param == "linear_features",
"train_means": param == "means",
"regularize_features": {"strength": 0.4, "norm": norm},
"regularize_predictors": {"strength": 0.5, "norm": norm},
"regularize_weights": {"strength": 0.7, "norm": norm},
"regularize_linear_features": {"strength": 0.3, "norm": norm},
"regularize_means": {"strength": 0.6, "norm": norm},
},
)
self.assertLess(err, 1e-6)
def test_gradient(self):
mcgsm = MCGSM(5, 2, 2, 4, 10)
cholesky_factors = []
for k in range(mcgsm.num_components):
cholesky_factors.append(cholesky(cov(randn(mcgsm.dim_out, mcgsm.dim_out**2))))
mcgsm.cholesky_factors = cholesky_factors
mcgsm.linear_features = randn(mcgsm.num_components, mcgsm.dim_in) / 5.
mcgsm.means = randn(mcgsm.dim_out, mcgsm.num_components) / 5.
err = mcgsm._check_gradient(
randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000), 1e-5)
self.assertLess(err, 1e-8)
# without regularization
for param in ['priors', 'scales', 'weights', 'features', 'chol', 'pred', 'linear_features', 'means']:
err = mcgsm._check_gradient(
randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000),
1e-5,
parameters={
'train_prior': param == 'priors',
'train_scales': param == 'scales',
'train_weights': param == 'weights',
'train_features': param == 'features',
'train_cholesky_factors': param == 'chol',
'train_predictors': param == 'pred',
'train_linear_features': param == 'linear_features',
'train_means': param == 'means',
})
self.assertLess(err, 1e-8)
# with regularization
for norm in ['L1', 'L2']:
for param in ['priors', 'scales', 'weights', 'features', 'chol', 'pred', 'linear_features', 'means']:
err = mcgsm._check_gradient(
randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000),
1e-7,
parameters={
'train_prior': param == 'priors',
'train_scales': param == 'scales',
'train_weights': param == 'weights',
'train_features': param == 'features',
'train_cholesky_factors': param == 'chol',
'train_predictors': param == 'pred',
'train_linear_features': param == 'linear_features',
'train_means': param == 'means',
'regularize_features': {'strength': 0.4, 'norm': norm},
'regularize_predictors': {'strength': 0.5, 'norm': norm},
'regularize_weights': {'strength': 0.7, 'norm': norm},
'regularize_linear_features': {'strength': 0.3, 'norm': norm},
'regularize_means': {'strength': 0.6, 'norm': norm},
})
self.assertLess(err, 1e-6)
def test_pickle(self):
mcgsm0 = MCGSM(11, 2, 4, 7, 21)
mcgsm0.linear_features = randn(mcgsm0.num_components, mcgsm0.dim_in)
mcgsm0.means = randn(mcgsm0.dim_out, mcgsm0.num_components)
tmp_file = mkstemp()[1]
# store model
with open(tmp_file, 'w') as handle:
dump({'mcgsm': mcgsm0}, handle)
# load model
with open(tmp_file) as handle:
mcgsm1 = load(handle)['mcgsm']
# make sure parameters haven't changed
self.assertEqual(mcgsm0.dim_in, mcgsm1.dim_in)
self.assertEqual(mcgsm0.dim_out, mcgsm1.dim_out)
self.assertEqual(mcgsm0.num_components, mcgsm1.num_components)
self.assertEqual(mcgsm0.num_scales, mcgsm1.num_scales)
self.assertEqual(mcgsm0.num_features, mcgsm1.num_features)
self.assertLess(max(abs(mcgsm0.scales - mcgsm1.scales)), 1e-20)
self.assertLess(max(abs(mcgsm0.weights - mcgsm1.weights)), 1e-20)
self.assertLess(max(abs(mcgsm0.features - mcgsm1.features)), 1e-20)
self.assertLess(
max(abs(mcgsm0.linear_features - mcgsm1.linear_features)), 1e-20)
self.assertLess(max(abs(mcgsm0.means - mcgsm1.means)), 1e-20)
for chol0, chol1 in zip(mcgsm0.cholesky_factors,
mcgsm1.cholesky_factors):
self.assertLess(max(abs(chol0 - chol1)), 1e-20)
for pred0, pred1 in zip(mcgsm0.predictors, mcgsm1.predictors):
self.assertLess(max(abs(pred0 - pred1)), 1e-20)
def test_gradient(self):
mcgsm = MCGSM(5, 2, 2, 4, 10)
cholesky_factors = []
for k in range(mcgsm.num_components):
cholesky_factors.append(
cholesky(cov(randn(mcgsm.dim_out, mcgsm.dim_out**2))))
mcgsm.cholesky_factors = cholesky_factors
mcgsm.linear_features = randn(mcgsm.num_components, mcgsm.dim_in) / 5.
mcgsm.means = randn(mcgsm.dim_out, mcgsm.num_components) / 5.
err = mcgsm._check_gradient(randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000), 1e-5)
self.assertLess(err, 1e-8)
# without regularization
for param in [
'priors', 'scales', 'weights', 'features', 'chol', 'pred',
'linear_features', 'means'
]:
err = mcgsm._check_gradient(
randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000),
1e-5,
parameters={
'train_prior': param == 'priors',
'train_scales': param == 'scales',
'train_weights': param == 'weights',
'train_features': param == 'features',
'train_cholesky_factors': param == 'chol',
'train_predictors': param == 'pred',
'train_linear_features': param == 'linear_features',
'train_means': param == 'means',
})
self.assertLess(err, 1e-8)
# with regularization
for norm in ['L1', 'L2']:
for param in [
'priors', 'scales', 'weights', 'features', 'chol', 'pred',
'linear_features', 'means'
]:
err = mcgsm._check_gradient(
randn(mcgsm.dim_in, 1000),
randn(mcgsm.dim_out, 1000),
1e-7,
parameters={
'train_prior': param == 'priors',
'train_scales': param == 'scales',
'train_weights': param == 'weights',
'train_features': param == 'features',
'train_cholesky_factors': param == 'chol',
'train_predictors': param == 'pred',
'train_linear_features': param == 'linear_features',
'train_means': param == 'means',
'regularize_features': {
'strength': 0.4,
'norm': norm
},
'regularize_predictors': {
'strength': 0.5,
'norm': norm
},
'regularize_weights': {
'strength': 0.7,
'norm': norm
},
'regularize_linear_features': {
'strength': 0.3,
'norm': norm
},
'regularize_means': {
'strength': 0.6,
'norm': norm
},
})
self.assertLess(err, 1e-6)
