def _test_gmm_parameters(covariance_type):
n_samples = [1000]
n_centers = [2]
stds = [.1, .5]
n_features = [2, 4]
for n in n_samples:
for f in n_features:
for c in n_centers:
for s in stds:
features, targets = generate_cluster_data(n_samples=n,
n_features=f,
n_centers=c,
cluster_stds=s)
# make model and fit
model = GMM(c, covariance_type=covariance_type)
model.fit(features)
covariances = model.covariances
for cov in covariances:
assert (np.abs(np.sqrt(cov) - s).mean() < 1e-1)
means = model.means
orderings = permutations(means)
distance_to_true_means = []
actual_means = np.array([
features[targets == i, :].mean(axis=0)
for i in range(targets.max() + 1)
])
for ordering in orderings:
_means = np.array(list(ordering))
distance_to_true_means.append(
np.abs(_means - actual_means).sum())
assert (min(distance_to_true_means) < 1e-1)
mixing_weights = model.mixing_weights
orderings = permutations(mixing_weights)
distance_to_true_mixing_weights = []
actual_mixing_weights = np.array([
features[targets == i, :].shape[0]
for i in range(targets.max() + 1)
])
actual_mixing_weights = actual_mixing_weights / actual_mixing_weights.sum(
)
for ordering in orderings:
_mixing_weights = np.array(list(ordering))
distance_to_true_mixing_weights.append(
np.abs(_mixing_weights -
actual_mixing_weights).sum())
assert (min(distance_to_true_mixing_weights) < 1e-1)
# predict and calculate adjusted mutual info
labels = model.predict(features)
acc = adjusted_mutual_info(targets, labels)
assert (acc >= .9)
def test_kmeans_spec():
features, targets = generate_cluster_data(n_samples=100,
n_features=2,
n_centers=2,
cluster_stds=.1)
model = KMeans(2)
model.fit(features)
assert (hasattr(model, 'means'))
def test_kmeans_spec():
features, targets = generate_cluster_data(n_samples=100,
n_features=2,
n_centers=2,
cluster_stds=.1)
model = KMeans(2)
rng_state = np.random.get_state()
model.fit(features)
np.random.set_state(rng_state)
assert (hasattr(model, 'means'))
def test_gmm_spec():
features, targets = generate_cluster_data(n_samples=100,
n_features=2,
n_centers=2,
cluster_stds=.1)
gmm = GMM(2, 'spherical')
gmm.fit(features)
assert (hasattr(gmm, 'means'))
assert (hasattr(gmm, 'covariances'))
assert (hasattr(gmm, 'mixing_weights'))
def test_kmeans_on_generated():
n_samples = [1000, 10000]
n_centers = [2]
stds = [.1]
n_features = [1, 2, 4]
for n in n_samples:
for f in n_features:
for c in n_centers:
for s in stds:
features, targets = generate_cluster_data(n_samples=n,
n_features=f,
n_centers=c,
cluster_stds=s)
# make model and fit
model = KMeans(c)
# Depending on how many random() calls the student code
# makes, it can mess with the random state used to generate
# data for subsequent tests and lead to an "impossible"
# input distribution that can't achieve the desired
# performance. To avoid this, we save and restore the
# random state so the student code can't interfere with it.
rng_state = np.random.get_state()
model.fit(features)
np.random.set_state(rng_state)
means = model.means
orderings = permutations(means)
distance_to_true_means = []
actual_means = np.array([
features[targets == i, :].mean(axis=0)
for i in range(targets.max() + 1)
])
for ordering in orderings:
_means = np.array(list(ordering))
distance_to_true_means.append(
np.abs(_means - actual_means).sum())
assert (min(distance_to_true_means) < 1e-1)
# predict and calculate adjusted mutual info
labels = model.predict(features)
acc = adjusted_mutual_info(targets, labels)
assert (acc >= .9)
def test_generate_cluster_data():
n_samples = [2000, 20000]
n_centers = [1, 2]
stds = [.1, .5, 1.0, 2.0]
n_features = [1, 2, 4]
for n in n_samples:
for f in n_features:
for c in n_centers:
for s in stds:
X, y = generate_cluster_data(n_samples=n,
n_features=f,
n_centers=c,
cluster_stds=s)
assert (X.shape == (n, f))
assert (y.max() == c - 1)
for i in range(y.max()):
subset = X[y == i]
assert (np.abs(np.std(subset, axis=0) - s).mean() < s)
def test_kmeans_on_generated():
n_samples = [1000, 10000]
n_centers = [2]
stds = [.1]
n_features = [1, 2, 4]
for n in n_samples:
for f in n_features:
for c in n_centers:
for s in stds:
features, targets = generate_cluster_data(n_samples=n,
n_features=f,
n_centers=c,
cluster_stds=s)
# make model and fit
model = KMeans(c)
model.fit(features)
means = model.means
orderings = permutations(means)
distance_to_true_means = []
actual_means = np.array([
features[targets == i, :].mean(axis=0)
for i in range(targets.max() + 1)
])
for ordering in orderings:
_means = np.array(list(ordering))
distance_to_true_means.append(
np.abs(_means - actual_means).sum())
assert (min(distance_to_true_means) < 1e-1)
# predict and calculate adjusted mutual info
labels = model.predict(features)
acc = adjusted_mutual_info(targets, labels)
assert (acc >= .9)
def _test_gmm_parameters(covariance_type):
n_samples = [1000]
n_centers = [2]
stds = [.1, .5]
n_features = [2, 4]
for n in n_samples:
for f in n_features:
for c in n_centers:
for s in stds:
features, targets = generate_cluster_data(n_samples=n,
n_features=f,
n_centers=c,
cluster_stds=s)
# make model and fit
model = GMM(c, covariance_type=covariance_type)
model.fit(features)
covariances = model.covariances
for cov in covariances:
print("mean cov: ", np.abs(np.sqrt(cov) - s).mean())
if (np.abs(np.sqrt(cov) - s).mean() < 1e-1):
return