def test_predict_non_gaussian(n_samples, n_features, n_neighbors, n_query):
np.random.seed(123)
X_host_train = pd.DataFrame(
np.random.uniform(0, 1, (n_samples, n_features)))
y_host_train = pd.DataFrame(np.random.randint(0, 5, (n_samples, 1)))
X_host_test = pd.DataFrame(np.random.uniform(0, 1, (n_query, n_features)))
X_device_train = cudf.DataFrame.from_pandas(X_host_train)
y_device_train = cudf.DataFrame.from_pandas(y_host_train)
X_device_test = cudf.DataFrame.from_pandas(X_host_test)
knn_sk = skKNN(algorithm="brute", n_neighbors=n_neighbors, n_jobs=1)
knn_sk.fit(X_host_train, y_host_train)
sk_result = knn_sk.predict(X_host_test)
knn_cuml = cuKNN(n_neighbors=n_neighbors)
knn_cuml.fit(X_device_train, y_device_train)
cuml_result = knn_cuml.predict(X_device_test)
assert np.array_equal(np.asarray(cuml_result.to_gpu_array()), sk_result)
def test_knn_x_none(input_type, nrows, n_feats, k, metric):
X, _ = make_blobs(n_samples=nrows,
n_features=n_feats, random_state=0)
p = 5 # Testing 5-norm of the minkowski metric only
knn_sk = skKNN(metric=metric, p=p) # Testing
knn_sk.fit(X.get())
D_sk, I_sk = knn_sk.kneighbors(X=None, n_neighbors=k)
X_orig = X
if input_type == "dataframe":
X = cudf.DataFrame(X)
knn_cu = cuKNN(metric=metric, p=p, output_type="numpy")
knn_cu.fit(X)
D_cuml, I_cuml = knn_cu.kneighbors(X=None, n_neighbors=k)
# Assert the cuml model was properly reverted
cp.testing.assert_allclose(knn_cu.X_m, X_orig,
atol=1e-5, rtol=1e-4)
# Allow a max relative diff of 10% and absolute diff of 1%
cp.testing.assert_allclose(D_cuml, D_sk, atol=5e-2,
rtol=1e-1)
assert I_cuml.all() == I_sk.all()
def test_predict_non_gaussian(n_samples, n_features, n_neighbors, n_query):
np.random.seed(123)
X_host_train = pd.DataFrame(
np.random.uniform(0, 1, (n_samples, n_features)))
y_host_train = pd.DataFrame(np.random.randint(0, 5, (n_samples, 1)))
X_host_test = pd.DataFrame(np.random.uniform(0, 1, (n_query, n_features)))
X_device_train = cudf.DataFrame.from_pandas(X_host_train)
y_device_train = cudf.DataFrame.from_pandas(y_host_train)
X_device_test = cudf.DataFrame.from_pandas(X_host_test)
knn_sk = skKNN(algorithm="brute", n_neighbors=n_neighbors, n_jobs=1)
knn_sk.fit(X_host_train, y_host_train.values.ravel())
sk_result = knn_sk.predict(X_host_test)
knn_cuml = cuKNN(n_neighbors=n_neighbors)
knn_cuml.fit(X_device_train, y_device_train)
with cuml.using_output_type("numpy"):
cuml_result = knn_cuml.predict(X_device_test)
assert np.array_equal(cuml_result, sk_result)
def test_cuml_against_sklearn(input_type, nrows, n_feats, k):
X, _ = make_blobs(n_samples=nrows,
n_features=n_feats, random_state=0)
knn_sk = skKNN(metric="euclidean")
knn_sk.fit(X)
D_sk, I_sk = knn_sk.kneighbors(X, k)
if input_type == "dataframe":
X = cudf.DataFrame.from_gpu_matrix(rmm.to_device(X))
knn_cu = cuKNN()
knn_cu.fit(X)
D_cuml, I_cuml = knn_cu.kneighbors(X, k)
if input_type == "dataframe":
assert isinstance(D_cuml, cudf.DataFrame)
assert isinstance(I_cuml, cudf.DataFrame)
D_cuml_arr = D_cuml.as_gpu_matrix().copy_to_host()
I_cuml_arr = I_cuml.as_gpu_matrix().copy_to_host()
else:
assert isinstance(D_cuml, np.ndarray)
assert isinstance(I_cuml, np.ndarray)
D_cuml_arr = D_cuml
I_cuml_arr = I_cuml
assert array_equal(D_cuml_arr, D_sk, 1e-2, with_sign=True)
assert I_cuml_arr.all() == I_sk.all()
def test_tsne_knn_graph_used(name, type_knn_graph):
datasets
X = eval("datasets.load_{}".format(name))().data
neigh = skKNN(n_neighbors=90) if type_knn_graph == 'sklearn' \
else cuKNN(n_neighbors=90)
neigh.fit(X)
knn_graph = neigh.kneighbors_graph(X, mode="distance")
tsne = TSNE()
# Perform tsne with normal knn_graph
Y = tsne.fit_transform(X, True, knn_graph)
trust_normal = trustworthiness(X, Y)
print("Trust = ", trust_normal)
X_garbage = np.ones(X.shape)
knn_graph_garbage = neigh.kneighbors_graph(X_garbage, mode="distance")
# Perform tsne with garbage knn_graph
Y = tsne.fit_transform(X, True, knn_graph_garbage)
trust_garbage = trustworthiness(X, Y)
print("Trust = ", trust_garbage)
assert (trust_normal - trust_garbage) > 0.15
Y = tsne.fit_transform(X, True, knn_graph_garbage.tocoo())
trust_garbage = trustworthiness(X, Y)
print("Trust = ", trust_garbage)
assert (trust_normal - trust_garbage) > 0.15
Y = tsne.fit_transform(X, True, knn_graph_garbage.tocsc())
trust_garbage = trustworthiness(X, Y)
print("Trust = ", trust_garbage)
assert (trust_normal - trust_garbage) > 0.15
def test_knn_separate_index_search(input_type, nrows, n_feats, k, metric):
X, _ = make_blobs(n_samples=nrows, n_features=n_feats, random_state=0)
X_index = X[:100]
X_search = X[101:]
p = 5 # Testing 5-norm of the minkowski metric only
knn_sk = skKNN(metric=metric, p=p) # Testing
knn_sk.fit(X_index.get())
D_sk, I_sk = knn_sk.kneighbors(X_search.get(), k)
X_orig = X_index
if input_type == "dataframe":
X_index = cudf.DataFrame(X_index)
X_search = cudf.DataFrame(X_search)
knn_cu = cuKNN(metric=metric, p=p)
knn_cu.fit(X_index)
D_cuml, I_cuml = knn_cu.kneighbors(X_search, k)
if input_type == "dataframe":
assert isinstance(D_cuml, cudf.DataFrame)
assert isinstance(I_cuml, cudf.DataFrame)
D_cuml_np = D_cuml.to_numpy()
I_cuml_np = I_cuml.to_numpy()
else:
assert isinstance(D_cuml, cp.ndarray)
assert isinstance(I_cuml, cp.ndarray)
D_cuml_np = D_cuml.get()
I_cuml_np = I_cuml.get()
with cuml.using_output_type("numpy"):
# Assert the cuml model was properly reverted
np.testing.assert_allclose(knn_cu.X_m,
X_orig.get(),
atol=1e-3,
rtol=1e-3)
if metric == 'braycurtis':
diff = D_cuml_np - D_sk
# Braycurtis has a few differences, but this is computed by FAISS.
# So long as the indices all match below, the small discrepancy
# should be okay.
assert len(diff[diff > 1e-2]) / X_search.shape[0] < 0.06
else:
np.testing.assert_allclose(D_cuml_np, D_sk, atol=1e-3, rtol=1e-3)
assert I_cuml_np.all() == I_sk.all()
def test_nearest_neighbors_sparse(shape,
metric,
n_neighbors,
batch_size_index,
batch_size_query):
nrows, ncols, density = shape
if nrows == 1 and n_neighbors > 1:
return
a = cp.sparse.random(nrows, ncols, format='csr', density=density,
random_state=35)
b = cp.sparse.random(nrows, ncols, format='csr', density=density,
random_state=38)
if metric == 'jaccard':
a = a.astype('bool').astype('float32')
b = b.astype('bool').astype('float32')
logger.set_level(logger.level_debug)
nn = cuKNN(metric=metric, p=2.0, n_neighbors=n_neighbors,
algorithm="brute", output_type="numpy",
verbose=logger.level_debug,
algo_params={"batch_size_index": batch_size_index,
"batch_size_query": batch_size_query})
nn.fit(a)
cuD, cuI = nn.kneighbors(b)
if metric not in sklearn.neighbors.VALID_METRICS_SPARSE['brute']:
a = a.todense()
b = b.todense()
sknn = skKNN(metric=metric, p=2.0, n_neighbors=n_neighbors,
algorithm="brute", n_jobs=-1)
sk_X = a.get()
sknn.fit(sk_X)
skD, skI = sknn.kneighbors(b.get())
cp.testing.assert_allclose(cuD, skD, atol=1e-3, rtol=1e-3)
# Jaccard & Chebyshev have a high potential for mismatched indices
# due to duplicate distances. We can ignore the indices in this case.
if metric not in ['jaccard', 'chebyshev']:
cp.testing.assert_allclose(cuI, skI, atol=1e-4, rtol=1e-4)
def test_tsne_knn_parameters_sparse(type_knn_graph, input_type):
datasets
digits = datasets.load_digits()
neigh = skKNN(n_neighbors=90) if type_knn_graph == 'sklearn' \
else cuKNN(n_neighbors=90)
digits_selection = np.random.RandomState(42).choice(
[True, False], 1797, replace=True, p=[0.60, 0.40])
selected_digits = digits.data[~digits_selection]
neigh.fit(selected_digits)
knn_graph = neigh.kneighbors_graph(selected_digits, mode="distance")
if input_type == 'cupy':
sp_prefix = cupyx.scipy.sparse
else:
sp_prefix = scipy.sparse
tsne = TSNE(2, n_neighbors=15,
random_state=1,
learning_rate=500,
angle=0.8)
new_data = sp_prefix.csr_matrix(
scipy.sparse.csr_matrix(selected_digits))
Y = tsne.fit_transform(new_data, True, knn_graph)
if input_type == 'cupy':
Y = Y.get()
check_embedding(selected_digits, Y, 0.85)
Y = tsne.fit_transform(new_data, True, knn_graph.tocoo())
if input_type == 'cupy':
Y = Y.get()
check_embedding(selected_digits, Y, 0.85)
Y = tsne.fit_transform(new_data, True, knn_graph.tocsc())
if input_type == 'cupy':
Y = Y.get()
check_embedding(selected_digits, Y, 0.85)
del Y
def test_ann_distances_metrics(algo, metric):
X, y = make_blobs(n_samples=500, centers=2, n_features=128, random_state=0)
cu_knn = cuKNN(algorithm=algo, metric=metric)
cu_knn.fit(X)
cu_dist, cu_ind = cu_knn.kneighbors(X,
n_neighbors=10,
return_distance=True)
del cu_knn
gc.collect()
X = X.get()
sk_knn = skKNN(metric=metric)
sk_knn.fit(X)
sk_dist, sk_ind = sk_knn.kneighbors(X,
n_neighbors=10,
return_distance=True)
return array_equal(sk_dist, cu_dist)
def test_knn_graph(input_type, mode, output_type, as_instance, nrows, n_feats,
p, k, metric):
X, _ = make_blobs(n_samples=nrows, n_features=n_feats, random_state=0)
if as_instance:
sparse_sk = sklearn.neighbors.kneighbors_graph(X.get(),
k,
mode=mode,
metric=metric,
p=p,
include_self='auto')
else:
knn_sk = skKNN(metric=metric, p=p)
knn_sk.fit(X.get())
sparse_sk = knn_sk.kneighbors_graph(X.get(), k, mode=mode)
if input_type == "dataframe":
X = cudf.DataFrame(X)
with cuml.using_output_type(output_type):
if as_instance:
sparse_cu = cuml.neighbors.kneighbors_graph(X,
k,
mode=mode,
metric=metric,
p=p,
include_self='auto')
else:
knn_cu = cuKNN(metric=metric, p=p)
knn_cu.fit(X)
sparse_cu = knn_cu.kneighbors_graph(X, k, mode=mode)
assert np.array_equal(sparse_sk.data.shape, sparse_cu.data.shape)
assert np.array_equal(sparse_sk.indices.shape, sparse_cu.indices.shape)
assert np.array_equal(sparse_sk.indptr.shape, sparse_cu.indptr.shape)
assert np.array_equal(sparse_sk.toarray().shape, sparse_cu.toarray().shape)
if output_type == 'cupy' or output_type is None:
assert cupyx.scipy.sparse.isspmatrix_csr(sparse_cu)
else:
assert isspmatrix_csr(sparse_cu)
def test_knn_search(input_type, should_downcast):
dtype = np.float32 if not should_downcast else np.float64
X = np.array(
[[1.0], [50.0], [51.0]],
dtype=dtype) # For now, FAISS only seems to support single precision
knn_sk = skKNN(X, metric="l2")
D_sk, I_sk = knn_sk.query(X, len(X))
knn_cu = cuKNN(should_downcast=should_downcast)
if input_type == 'dataframe':
X = cudf.DataFrame.from_pandas(pd.DataFrame(X))
knn_cu.fit(X)
D_cuml, I_cuml = knn_cu.query(X, len(X))
assert type(D_cuml) == cudf.DataFrame
assert type(I_cuml) == cudf.DataFrame
# FAISS does not perform sqrt on L2 because it's expensive
D_cuml_arr = np.asarray(D_cuml.as_gpu_matrix(order="C"))
I_cuml_arr = np.asarray(I_cuml.as_gpu_matrix(order="C"))
else:
knn_cu.fit(X)
D_cuml, I_cuml = knn_cu.query(X, len(X))
assert type(D_cuml) == np.ndarray
assert type(I_cuml) == np.ndarray
D_cuml_arr = D_cuml
I_cuml_arr = I_cuml
print(str(D_cuml_arr))
print(str(D_cuml_arr))
print(str(I_cuml_arr))
assert np.array_equal(D_cuml_arr, np.square(D_sk))
assert np.array_equal(I_cuml_arr, I_sk)
def test_knn(input_type, should_downcast, nrows, n_feats, k):
n_samples = nrows
X, y = make_blobs(n_samples=n_samples,
n_features=n_feats, random_state=0)
knn_cu = cuKNN(should_downcast=should_downcast)
if input_type == 'dataframe':
X_pd = pd.DataFrame({'fea%d' % i: X[0:, i] for i in range(X.shape[1])})
X_cudf = cudf.DataFrame.from_pandas(X_pd)
knn_cu.fit(X_cudf)
D_cuml, I_cuml = knn_cu.kneighbors(X_cudf, k)
assert type(D_cuml) == cudf.DataFrame
assert type(I_cuml) == cudf.DataFrame
# FAISS does not perform sqrt on L2 because it's expensive
D_cuml_arr = np.asarray(D_cuml.as_gpu_matrix(order="C"))
I_cuml_arr = np.asarray(I_cuml.as_gpu_matrix(order="C"))
elif input_type == 'ndarray':
knn_cu.fit(X)
D_cuml, I_cuml = knn_cu.kneighbors(X, k)
assert type(D_cuml) == np.ndarray
assert type(I_cuml) == np.ndarray
D_cuml_arr = D_cuml
I_cuml_arr = I_cuml
if nrows < 500000:
knn_sk = skKNN(metric="l2")
knn_sk.fit(X)
D_sk, I_sk = knn_sk.kneighbors(X, k)
assert array_equal(D_cuml_arr, np.square(D_sk), 1e-2, with_sign=True)
assert I_cuml_arr.all() == I_sk.all()
def test_tsne_knn_parameters(name, type_knn_graph):
datasets
X = eval("datasets.load_{}".format(name))().data
neigh = skKNN(n_neighbors=90) if type_knn_graph == 'sklearn' \
else cuKNN(n_neighbors=90)
neigh.fit(X)
knn_graph = neigh.kneighbors_graph(X, mode="distance")
for i in range(3):
print("iteration = ", i)
tsne = TSNE()
Y = tsne.fit_transform(X, True, knn_graph)
check_embedding(X, Y)
Y = tsne.fit_transform(X, True, knn_graph.tocoo())
check_embedding(X, Y)
Y = tsne.fit_transform(X, True, knn_graph.tocsc())
check_embedding(X, Y)
del Y
def test_knn(input_type, nrows, n_feats, k, metric):
X, _ = make_blobs(n_samples=nrows, n_features=n_feats, random_state=0)
p = 5 # Testing 5-norm of the minkowski metric only
knn_sk = skKNN(metric=metric, p=p) # Testing
knn_sk.fit(X)
D_sk, I_sk = knn_sk.kneighbors(X, k)
X_orig = X
if input_type == "dataframe":
X = cudf.DataFrame(X)
knn_cu = cuKNN(metric=metric, p=p)
knn_cu.fit(X)
D_cuml, I_cuml = knn_cu.kneighbors(X, k)
if input_type == "dataframe":
assert isinstance(D_cuml, cudf.DataFrame)
assert isinstance(I_cuml, cudf.DataFrame)
D_cuml_arr = D_cuml.as_gpu_matrix().copy_to_host()
I_cuml_arr = I_cuml.as_gpu_matrix().copy_to_host()
else:
assert isinstance(D_cuml, np.ndarray)
assert isinstance(I_cuml, np.ndarray)
D_cuml_arr = D_cuml
I_cuml_arr = I_cuml
# Assert the cuml model was properly reverted
np.testing.assert_allclose(knn_cu._X_m.to_output("numpy"),
X_orig,
atol=1e-5,
rtol=1e-4)
# Allow a max relative diff of 10% and absolute diff of 1%
np.testing.assert_allclose(D_cuml_arr, D_sk, atol=1e-2, rtol=1e-1)
assert I_cuml_arr.all() == I_sk.all()
def test_nearest_neighbors_sparse(nrows, ncols, density, metric, n_neighbors,
batch_size_index, batch_size_query):
if nrows == 1 and n_neighbors > 1:
return
a = cp.sparse.random(nrows,
ncols,
format='csr',
density=density,
random_state=32)
logger.set_level(logger.level_info)
nn = cuKNN(metric=metric,
n_neighbors=n_neighbors,
algorithm="brute",
verbose=logger.level_debug,
algo_params={
"batch_size_index": batch_size_index,
"batch_size_query": batch_size_query
})
nn.fit(a)
cuD, cuI = nn.kneighbors(a)
sknn = skKNN(metric=metric,
n_neighbors=n_neighbors,
algorithm="brute",
n_jobs=-1)
sk_X = a.get()
sknn.fit(sk_X)
skD, skI = sknn.kneighbors(sk_X)
cp.testing.assert_allclose(cuI, skI, atol=1e-4, rtol=1e-4)
cp.testing.assert_allclose(cuD, skD, atol=1e-3, rtol=1e-3)
def test_nearest_neighbors_sparse(metric, nrows, ncols, density, n_neighbors,
batch_size_index, batch_size_query):
if nrows == 1 and n_neighbors > 1:
return
a = cupyx.scipy.sparse.random(nrows,
ncols,
format='csr',
density=density,
random_state=35)
b = cupyx.scipy.sparse.random(nrows,
ncols,
format='csr',
density=density,
random_state=38)
if metric == 'jaccard':
a = a.astype('bool').astype('float32')
b = b.astype('bool').astype('float32')
logger.set_level(logger.level_debug)
nn = cuKNN(metric=metric,
p=2.0,
n_neighbors=n_neighbors,
algorithm="brute",
output_type="numpy",
verbose=logger.level_debug,
algo_params={
"batch_size_index": batch_size_index,
"batch_size_query": batch_size_query
})
nn.fit(a)
cuD, cuI = nn.kneighbors(b)
if metric not in sklearn.neighbors.VALID_METRICS_SPARSE['brute']:
a = a.todense()
b = b.todense()
sknn = skKNN(metric=metric,
p=2.0,
n_neighbors=n_neighbors,
algorithm="brute",
n_jobs=-1)
sk_X = a.get()
sknn.fit(sk_X)
skD, skI = sknn.kneighbors(b.get())
# For some reason, this will occasionally fail w/ a single
# mismatched element in CI. Allowing the single mismatch for now.
cp.testing.assert_allclose(cuD, skD, atol=1e-5, rtol=1e-5)
# Jaccard & Chebyshev have a high potential for mismatched indices
# due to duplicate distances. We can ignore the indices in this case.
if metric not in ['jaccard', 'chebyshev']:
# The actual neighbors returned in the presence of duplicate distances
# is non-deterministic. If we got to this point, the distances all
# match between cuml and sklearn. We set a reasonable threshold
# (.5% in this case) to allow differences from non-determinism.
diffs = abs(cuI - skI)
assert (len(diffs[diffs > 0]) / len(np.ravel(skI))) <= 0.005