Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add ids to Python FeatureVectorArray #442

Merged
merged 3 commits into from
Jul 12, 2024
Merged

Add ids to Python FeatureVectorArray #442

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0cdd4a1
add ids to python FeatureVectorArray
jparismorgan Jul 11, 2024
1cbab1f
cleanup and more tests
jparismorgan Jul 11, 2024
214c99e
fix build
jparismorgan Jul 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
158 changes: 120 additions & 38 deletions apis/python/src/tiledb/vector_search/type_erased_module.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,10 +97,36 @@ auto datatype_to_format(tiledb_datatype_t datatype) {
case TILEDB_UINT64:
return py::format_descriptor<uint64_t>::format();
default:
throw std::runtime_error("Unsupported datatype");
throw std::runtime_error(
"[type_erased_module@datatype_to_format] Unsupported datatype");
}
}

bool check_datatype_format(
const std::string& dtype_format, const std::string& buffer_info_format) {
if (dtype_format == buffer_info_format) {
return true;
}
// We need to handle uint64 specifically of a numpy quirk:
// a. dtype_format (i.e.
// `datatype_to_format(string_to_datatype(<py::array>.dtype().str()))`) will
// give us 'Q' (numpy.ulonglong) See:
// https://numpy.org/doc/stable/reference/arrays.scalars.html#numpy.ulonglong
// b. buffer_info_format (i.e. `<py::array>.request().format`) will
// give us 'L' (numpy.uint) because numpy.uint is an alias for numpy.uint64 on
// Darwin arm64. See:
// https://numpy.org/doc/stable/reference/arrays.scalars.html#numpy.uint
if (dtype_format == "Q" && buffer_info_format == "L") {
return true;
}
// The same thing happens with int64, but for it dtype_format will give 'q'
// (numpy.longlong), whereas buffer_info_format gives 'l' (numpy.int_).
if (dtype_format == "q" && buffer_info_format == "l") {
return true;
}
return false;
}

// Define Pybind11 bindings

// PYBIND11_MODULE(_tiledbvspy2, m) {
Expand Down Expand Up @@ -176,19 +202,26 @@ void init_type_erased_module(py::module_& m) {
/* Strides (in bytes) for each index */
);
})
.def(py::init([](py::array b) {
.def(py::init([](py::array vector) {
/* Request a buffer descriptor from Python */
py::buffer_info info = b.request();
if (info.ndim != 1)
py::buffer_info info = vector.request();
if (info.ndim != 1) {
throw std::runtime_error(
"Incompatible buffer dimension! Should be 1.");
"[type_erased_module@FeatureVector] Incompatible buffer "
"dimension. Should be 1, but was " +
std::to_string(info.ndim) + ".");
}

auto dtype_str = b.dtype().str();
auto dtype_str = vector.dtype().str();
tiledb_datatype_t datatype = string_to_datatype(dtype_str);
if (info.format != datatype_to_format(datatype))
auto datatype_format = datatype_to_format(datatype);
if (!check_datatype_format(datatype_format, info.format)) {
throw std::runtime_error(
"Incompatible format: expected array of " +
datatype_to_string(datatype));
"[type_erased_module@FeatureVector] Incompatible format: "
"expected array of " +
datatype_to_string(datatype) + " (" + datatype_format +
"), but was " + info.format + ".");
}

size_t sz = datatype_to_size(datatype);

Expand Down Expand Up @@ -243,38 +276,84 @@ void init_type_erased_module(py::module_& m) {
v.dimensions(), /* Strides (in bytes) for each index */
datatype_to_size(v.feature_type())});
})
.def(py::init([](py::array b) {
/* Request a buffer descriptor from Python */
py::buffer_info info = b.request();
if (info.ndim != 2)
throw std::runtime_error(
"Incompatible buffer dimension! Should be 2.");
.def(
py::init([](py::array vectors, py::array ids) {
// The vector buffer info.
py::buffer_info info = vectors.request();
if (info.ndim != 2) {
throw std::runtime_error(
"[type_erased_module@FeatureVectorArray] Incompatible buffer "
"dimension. Should be 2, but was " +
std::to_string(info.ndim) + ".");
}

auto dtype_str = b.dtype().str();
tiledb_datatype_t datatype = string_to_datatype(dtype_str);
if (info.format != datatype_to_format(datatype))
throw std::runtime_error(
"Incompatible format: expected array of " +
datatype_to_string(datatype));
auto dtype_str = vectors.dtype().str();
tiledb_datatype_t datatype = string_to_datatype(dtype_str);
auto datatype_format = datatype_to_format(datatype);
if (!check_datatype_format(datatype_format, info.format)) {
throw std::runtime_error(
"[type_erased_module@FeatureVectorArray] Incompatible format "
"- expected array of " +
datatype_to_string(datatype) + " (" + datatype_format +
"), but was " + info.format + ".");
}

size_t sz = datatype_to_size(datatype);
// The ids vector buffer info.
py::buffer_info ids_info = ids.request();
if (ids_info.ndim != 1) {
throw std::runtime_error(
"[type_erased_module@FeatureVectorArray] Incompatible ids "
"buffer dimension. Should be 1, but was " +
std::to_string(ids_info.ndim) + ".");
}

auto v = [&]() {
auto order = b.flags() & py::array::f_style ? TILEDB_COL_MAJOR :
TILEDB_ROW_MAJOR;
if (order == TILEDB_COL_MAJOR) {
return FeatureVectorArray(info.shape[0], info.shape[1], dtype_str);
} else {
return FeatureVectorArray(info.shape[1], info.shape[0], dtype_str);
}
}();
std::string ids_dtype_str;
tiledb_datatype_t ids_datatype = TILEDB_ANY;
if (ids.size() != 0) {
ids_dtype_str = ids.dtype().str();
ids_datatype = string_to_datatype(ids_dtype_str);
auto ids_datatype_format = datatype_to_format(ids_datatype);
if (!check_datatype_format(
ids_datatype_format, ids_info.format)) {
throw std::runtime_error(
"[type_erased_module@FeatureVectorArray] Incompatible ids "
"format - expected array of " +
datatype_to_string(ids_datatype) + " (" +
ids_datatype_format + "), but was " + ids_info.format +
".");
}
}

auto data = (uint8_t*)v.data();
std::memcpy(
data, (uint8_t*)info.ptr, info.shape[0] * info.shape[1] * sz);
auto feature_vector_array = [&]() {
auto order = vectors.flags() & py::array::f_style ?
TILEDB_COL_MAJOR :
TILEDB_ROW_MAJOR;
if (order == TILEDB_COL_MAJOR) {
return FeatureVectorArray(
info.shape[0], info.shape[1], dtype_str, ids_dtype_str);
} else {
return FeatureVectorArray(
info.shape[1], info.shape[0], dtype_str, ids_dtype_str);
}
}();

auto data = (uint8_t*)feature_vector_array.data();
std::memcpy(
data,
(uint8_t*)info.ptr,
info.shape[0] * info.shape[1] * datatype_to_size(datatype));

if (ids.size() != 0) {
std::memcpy(
feature_vector_array.ids(),
(uint8_t*)ids_info.ptr,
ids_info.shape[0] * datatype_to_size(ids_datatype));
}

return v;
}));
return feature_vector_array;
}),
py::arg("vectors"),
py::arg("ids") = py::array());

py::class_<IndexFlatL2>(m, "IndexFlatL2")
.def(
Expand All @@ -301,7 +380,8 @@ void init_type_erased_module(py::module_& m) {
} else if (s == "random") {
return kmeans_init::random;
} else {
throw std::runtime_error("Invalid kmeans_init value");
throw std::runtime_error(
"[type_erased_module@kmeans_init] Invalid kmeans_init value");
}
}));

Expand Down Expand Up @@ -478,7 +558,9 @@ void init_type_erased_module(py::module_& m) {
} else if (std::string(init_str) == "random") {
init = kmeans_init::random;
} else {
throw std::runtime_error("Invalid kmeans_init value");
throw std::runtime_error(
"[type_erased_module@IndexIVFFlat@train] Invalid kmeans_init "
"value");
}
index.train(vectors, init);
},
Expand Down
71 changes: 71 additions & 0 deletions apis/python/test/test_type_erased_module.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,30 @@ def test_feature_vector_to_numpy():
assert b.dtype == np.uint64


def test_numpy_to_feature_vector_data_types():
for dtype in [
np.float32,
np.int8,
np.uint8,
np.int32,
np.uint32,
np.uint64,
]:
if np.issubdtype(dtype, np.integer):
max_val = np.iinfo(dtype).max
elif np.issubdtype(dtype, np.floating):
max_val = np.finfo(dtype).max
else:
raise TypeError(f"Unsupported data type {dtype}")

vector = np.array([max_val], dtype=dtype)
feature_vector = vspy.FeatureVector(vector)
assert feature_vector.feature_type_string() == np.dtype(dtype).name
assert np.array_equal(
vector, np.array(feature_vector)
), f"Arrays were not equal for dtype: {dtype}"


def test_numpy_to_feature_vector_array_simple():
a = np.array(np.random.rand(10000), dtype=np.float32)
b = vspy.FeatureVector(a)
Expand Down Expand Up @@ -73,16 +97,55 @@ def test_feature_vector_array_to_numpy():
a = vspy.FeatureVectorArray(ctx, siftsmall_inputs_uri)
assert a.num_vectors() == 10000
assert a.dimensions() == 128
assert a.num_ids() == 0
assert a.ids_type_string() == "any"
b = np.array(a)
assert b.shape == (10000, 128)

a = vspy.FeatureVectorArray(ctx, bigann10k_inputs_uri)
assert a.num_vectors() == 10000
assert a.dimensions() == 128
assert a.num_ids() == 0
assert a.ids_type_string() == "any"
b = np.array(a)
assert b.shape == (10000, 128)


def test_numpy_to_feature_vector_array_data_types():
for dtype in [
np.float32,
np.int8,
np.uint8,
np.int32,
np.uint32,
np.int64,
np.uint64,
]:
for dtype_ids in [np.uint32, np.uint64]:
if np.issubdtype(dtype, np.integer):
max_val = np.iinfo(dtype).max
elif np.issubdtype(dtype, np.floating):
max_val = np.finfo(dtype).max
else:
raise TypeError(f"Unsupported data type {dtype}")

if np.issubdtype(dtype_ids, np.integer):
max_val_ids = np.iinfo(dtype_ids).max
elif np.issubdtype(dtype, np.floating):
max_val_ids = np.finfo(dtype_ids).max
else:
raise TypeError(f"Unsupported ids data type {dtype_ids}")

vectors = np.array([[max_val]], dtype=dtype)
ids = np.array([max_val_ids], dtype=dtype_ids)
feature_vector_array = vspy.FeatureVectorArray(vectors, ids)
assert feature_vector_array.feature_type_string() == np.dtype(dtype).name
assert feature_vector_array.ids_type_string() == np.dtype(dtype_ids).name
assert np.array_equal(
vectors, np.array(feature_vector_array)
), f"Arrays were not equal for dtype: {dtype}, dtype_ids: {dtype_ids}"


def test_numpy_to_feature_vector_array():
a = np.array(np.random.rand(10000, 128), dtype=np.float32)
b = vspy.FeatureVectorArray(a)
Expand Down Expand Up @@ -162,6 +225,14 @@ def test_numpy_to_feature_vector_array():
assert np.array_equal(a, np.transpose(np.array(b)))


def test_numpy_to_feature_vector_array_with_ids():
a = np.array(np.random.rand(10000, 128), dtype=np.float32)
ids = np.arange(10000, dtype=np.uint64)
b = vspy.FeatureVectorArray(a, ids)
assert b.num_ids() == 10000
assert b.ids_type_string() == "uint64"


def test_TemporalPolicy():
temporal_policy = vspy.TemporalPolicy()
assert temporal_policy.timestamp_start() == 0
Expand Down
Loading