def test_speed_vs_sk(self, order):
from h2o4gpu.cluster import KMeansSklearn as skKMeans
n_samples = 100000
centers = 10
X, true_labels = make_blobs(n_samples=n_samples,
centers=centers,
cluster_std=1.,
random_state=42)
kmeans_h2o = KMeans(n_gpus=1, n_clusters=centers, random_state=42)
# Warmup - during first call CUDA kernels take ~2sec to load
kmeans_h2o.fit(X)
start_h2o = time.time()
kmeans_h2o.fit(X)
end_h2o = time.time()
if os.getenv("CHECKPERFORMANCE") is not None:
kmeans_sk = skKMeans(n_init=1,
n_clusters=centers,
algorithm='full',
n_jobs=-1)
start_sk = time.time()
kmeans_sk.fit(X)
end_sk = time.time()
assert end_h2o - start_h2o <= end_sk - start_sk
def test_speed_vs_sk(self):
from sklearn.cluster import KMeans as skKMeans
n_samples = 100000
centers = 10
X, true_labels = make_blobs(n_samples=n_samples,
centers=centers,
cluster_std=1.,
random_state=42)
kmeans_h2o = KMeans(n_gpus=1, n_clusters=centers, random_state=42)
# Warmup - during first call CUDA kernels take ~2sec to load
kmeans_h2o.fit(X)
start_h2o = time.time()
kmeans_h2o.fit(X)
end_h2o = time.time()
kmeans_sk = skKMeans(n_init=1,
n_clusters=centers,
init='random',
algorithm='full',
n_jobs=-1)
start_sk = time.time()
kmeans_sk.fit(X)
end_sk = time.time()
assert end_h2o - start_h2o <= end_sk - start_sk
def _test_accuracy(self,
order,
n_samples=500000,
centers=10,
n_features=2):
from h2o4gpu.cluster import KMeansSklearn as skKMeans
X, true_labels = make_blobs(n_samples=n_samples,
centers=centers,
n_features=n_features,
cluster_std=2.,
random_state=42)
X = np.asanyarray(X, order=order)
kmeans_h2o = KMeans(n_gpus=1,
n_clusters=centers,
random_state=42,
verbose=1000)
kmeans_h2o.fit(X)
kmeans_sk = skKMeans(n_init=1, n_clusters=centers, random_state=42)
kmeans_sk.fit(X)
accuracy_h2o = v_measure_score(kmeans_h2o.labels_, true_labels)
accuracy_sk = v_measure_score(kmeans_sk.labels_, true_labels)
# We also want to be either better or at most 10% worse than SKLearn
# Everything else is horrible and we probably should fix something
# TODO: it's failing with lower rtol, find out why it's so inaccurate
assert np.allclose(accuracy_h2o, accuracy_sk, rtol=0.1, atol=0.05), \
'Accuracy error {0} {1} n_samples={2}, centers={3}, n_features={4}, order={5}'.format(
accuracy_h2o, accuracy_sk, n_samples, centers, n_features, order)
def test_fit_i32_vs_f32(self):
X_f64 = np.array([[1., 2.], [1., 4.], [1., 0.], [4., 2.], [4., 4.], [4., 0.]])
X_f32 = X_f64.astype(np.float32)
X_i32 = X_f64.astype(np.int32)
kmeans = KMeans(n_gpus=1, n_clusters=2, random_state=123)
model_f64_labels = kmeans.fit(X_f64).predict(X_f64)
model_f32_labels = kmeans.fit(X_f32).predict(X_f32)
model_i32_labels = kmeans.fit(X_i32).predict(X_i32)
assert all(model_f64_labels == model_f32_labels)
assert all(model_f32_labels == model_i32_labels)
def test_fit_iris(self):
X = load_iris().data
clusters = 4
model = KMeans(n_gpus=1, n_clusters=clusters, random_state=123).fit(X)
assert model.cluster_centers_.shape == (X.shape[1], clusters)
model_rerun = KMeans(n_gpus=1, n_clusters=clusters,
random_state=123).fit(X)
# Same random_state should yield same results
assert np.allclose(model.cluster_centers_,
model_rerun.cluster_centers_)
model_rerun2 = model_rerun.fit(X)
# Multiple invocations of fit with the same random_state
# also should produce the same result
assert np.allclose(model_rerun.cluster_centers_,
model_rerun2.cluster_centers_)
model_all = KMeans(n_clusters=clusters, random_state=123).fit(X)
# Multi GPU should yield same result as single GPU
assert np.allclose(model.cluster_centers_, model_all.cluster_centers_)
def test_fit_iris(self):
X = load_iris().data
clusters = 4
model = KMeans(n_gpus=1, n_clusters=clusters, random_state=123).fit(X)
assert model.cluster_centers_.shape == (X.shape[1], clusters)
model_rerun = KMeans(n_gpus=1, n_clusters=clusters, random_state=123).fit(X)
# Same random_state should yield same results
assert np.allclose(
model.cluster_centers_, model_rerun.cluster_centers_
)
model_rerun2 = model_rerun.fit(X)
# Multiple invocations of fit with the same random_state
# also should produce the same result
assert np.allclose(
model_rerun.cluster_centers_, model_rerun2.cluster_centers_
)
model_all = KMeans(n_clusters=clusters, random_state=123).fit(X)
# Multi GPU should yield same result as single GPU
assert np.allclose(
model.cluster_centers_, model_all.cluster_centers_
)
def test_fit_iris_precision(self):
X_f64 = load_iris().data
X_f32 = X_f64.astype(np.float32)
kmeans = KMeans(n_gpus=1, n_clusters=4, random_state=12345)
model_f64_labels = kmeans.fit(X_f64).predict(X_f64)
model_f32_labels = kmeans.fit(X_f32).predict(X_f32)
assert all(model_f64_labels == model_f32_labels)
def test_fit_iris_precision(self):
X_f64 = load_iris().data
X_f32 = X_f64.astype(np.float32)
kmeans = KMeans(n_gpus=1, n_clusters=4, random_state=12345)
model_f64_labels = kmeans.fit(X_f64).predict(X_f64)
model_f32_labels = kmeans.fit(X_f32).predict(X_f32)
assert all(model_f64_labels == model_f32_labels)
def test_accuracy(self):
from sklearn.cluster import KMeans as skKMeans
n_samples = 500000
centers = 10
X, true_labels = make_blobs(n_samples=n_samples, centers=centers,
cluster_std=1., random_state=42)
kmeans_h2o = KMeans(n_gpus=1, n_clusters=centers, random_state=42)
kmeans_h2o.fit(X)
kmeans_sk = skKMeans(n_init=1, n_clusters=centers, random_state=42)
kmeans_sk.fit(X)
accuracy_h2o = v_measure_score(kmeans_h2o.labels_, true_labels)
accuracy_sk = v_measure_score(kmeans_sk.labels_, true_labels)
# We also want to be either better or at most 10% worse than SKLearn
# Everything else is horrible and we probably should fix something
assert accuracy_h2o - accuracy_sk >= -0.1
def test_accuracy(self):
from sklearn.cluster import KMeans as skKMeans
n_samples = 100000
centers = 10
X, true_labels = make_blobs(n_samples=n_samples, centers=centers,
cluster_std=1., random_state=42)
kmeans_h2o = KMeans(n_gpus=1, n_clusters=centers, random_state=42)
kmeans_h2o.fit(X)
kmeans_sk = skKMeans(n_init=1, n_clusters=centers, init='random',
random_state=42)
kmeans_sk.fit(X)
accuracy_h2o = v_measure_score(kmeans_h2o.labels_, true_labels)
accuracy_sk = v_measure_score(kmeans_sk.labels_, true_labels)
# We also want to be either better or at most 10% worse than SKLearn
# Everything else is horrible and we probably should fix something
assert accuracy_h2o - accuracy_sk >= -0.1
def test_speed_vs_sk(self):
from sklearn.cluster import KMeans as skKMeans
n_samples = 100000
centers = 10
X, true_labels = make_blobs(n_samples=n_samples, centers=centers,
cluster_std=1., random_state=42)
kmeans_h2o = KMeans(n_gpus=1, n_clusters=centers)
start_h2o = time.time()
kmeans_h2o.fit(X)
end_h2o = time.time()
kmeans_sk = skKMeans(n_init=1, n_clusters=centers, init='random')
start_sk = time.time()
kmeans_sk.fit(X)
end_sk = time.time()
print(end_h2o - start_h2o)
print(end_sk - start_sk)