SungBeom's picture
Upload folder using huggingface_hub
4a51346
raw
history blame
11.1 kB
import os
import pickle
import time
from typing import Dict, List, Optional, Set, Tuple, Union, cast
from chromadb.api.types import Embeddings, IndexMetadata
import hnswlib
from chromadb.config import Settings
from chromadb.db.index import Index
from chromadb.errors import (
InvalidDimensionException,
)
import logging
import re
from uuid import UUID
import multiprocessing
logger = logging.getLogger(__name__)
valid_params = {
"hnsw:space": r"^(l2|cosine|ip)$",
"hnsw:construction_ef": r"^\d+$",
"hnsw:search_ef": r"^\d+$",
"hnsw:M": r"^\d+$",
"hnsw:num_threads": r"^\d+$",
"hnsw:resize_factor": r"^\d+(\.\d+)?$",
}
DEFAULT_CAPACITY = 1000
class HnswParams:
space: str
construction_ef: int
search_ef: int
M: int
num_threads: int
resize_factor: float
def __init__(self, metadata: Dict[str, str]):
metadata = metadata or {}
# Convert all values to strings for future compatibility.
metadata = {k: str(v) for k, v in metadata.items()}
for param, value in metadata.items():
if param.startswith("hnsw:"):
if param not in valid_params:
raise ValueError(f"Unknown HNSW parameter: {param}")
if not re.match(valid_params[param], value):
raise ValueError(
f"Invalid value for HNSW parameter: {param} = {value}"
)
self.space = metadata.get("hnsw:space", "l2")
self.construction_ef = int(metadata.get("hnsw:construction_ef", 100))
self.search_ef = int(metadata.get("hnsw:search_ef", 10))
self.M = int(metadata.get("hnsw:M", 16))
self.num_threads = int(
metadata.get("hnsw:num_threads", multiprocessing.cpu_count())
)
self.resize_factor = float(metadata.get("hnsw:resize_factor", 1.2))
def hexid(id: Union[str, UUID]) -> str:
"""Backwards compatibility for old indexes which called uuid.hex on UUID ids"""
return id.hex if isinstance(id, UUID) else id
def delete_all_indexes(settings: Settings) -> None:
if os.path.exists(f"{settings.persist_directory}/index"):
for file in os.listdir(f"{settings.persist_directory}/index"):
os.remove(f"{settings.persist_directory}/index/{file}")
class Hnswlib(Index):
_id: str
_index: hnswlib.Index
_index_metadata: IndexMetadata
_params: HnswParams
_id_to_label: Dict[str, int]
_label_to_id: Dict[int, UUID]
def __init__(
self,
id: str,
settings: Settings,
metadata: Dict[str, str],
number_elements: int,
):
self._save_folder = settings.persist_directory + "/index"
self._params = HnswParams(metadata)
self._id = id
self._index = None
# Mapping of IDs to HNSW integer labels
self._id_to_label = {}
self._label_to_id = {}
self._load(number_elements)
def _init_index(self, dimensionality: int) -> None:
# more comments available at the source: https://github.com/nmslib/hnswlib
index = hnswlib.Index(
space=self._params.space, dim=dimensionality
) # possible options are l2, cosine or ip
index.init_index(
max_elements=DEFAULT_CAPACITY,
ef_construction=self._params.construction_ef,
M=self._params.M,
)
index.set_ef(self._params.search_ef)
index.set_num_threads(self._params.num_threads)
self._index = index
self._index_metadata = {
"dimensionality": dimensionality,
"curr_elements": 0,
"total_elements_added": 0,
"time_created": time.time(),
}
self._save()
def _check_dimensionality(self, data: Embeddings) -> None:
"""Assert that the given data matches the index dimensionality"""
dim = len(data[0])
idx_dim = self._index.dim
if dim != idx_dim:
raise InvalidDimensionException(
f"Dimensionality of ({dim}) does not match index dimensionality ({idx_dim})"
)
def add(
self, ids: List[UUID], embeddings: Embeddings, update: bool = False
) -> None:
"""Add or update embeddings to the index"""
dim = len(embeddings[0])
if self._index is None:
self._init_index(dim)
# Calling init_index will ensure the index is not none, so we can safely cast
self._index = cast(hnswlib.Index, self._index)
# Check dimensionality
self._check_dimensionality(embeddings)
labels = []
for id in ids:
if hexid(id) in self._id_to_label:
if update:
labels.append(self._id_to_label[hexid(id)])
else:
raise ValueError(f"ID {id} already exists in index")
else:
self._index_metadata["total_elements_added"] += 1
self._index_metadata["curr_elements"] += 1
next_label = self._index_metadata["total_elements_added"]
self._id_to_label[hexid(id)] = next_label
self._label_to_id[next_label] = id
labels.append(next_label)
if (
self._index_metadata["total_elements_added"]
> self._index.get_max_elements()
):
new_size = int(
max(
self._index_metadata["total_elements_added"]
* self._params.resize_factor,
DEFAULT_CAPACITY,
)
)
self._index.resize_index(new_size)
self._index.add_items(embeddings, labels)
self._save()
def delete(self) -> None:
# delete files, dont throw error if they dont exist
try:
os.remove(f"{self._save_folder}/id_to_uuid_{self._id}.pkl")
os.remove(f"{self._save_folder}/uuid_to_id_{self._id}.pkl")
os.remove(f"{self._save_folder}/index_{self._id}.bin")
os.remove(f"{self._save_folder}/index_metadata_{self._id}.pkl")
except Exception:
pass
self._index = None
self._collection_uuid = None
self._id_to_label = {}
self._label_to_id = {}
def delete_from_index(self, ids: List[UUID]) -> None:
if self._index is not None:
for id in ids:
label = self._id_to_label[hexid(id)]
self._index.mark_deleted(label)
del self._label_to_id[label]
del self._id_to_label[hexid(id)]
self._index_metadata["curr_elements"] -= 1
self._save()
def _save(self) -> None:
# create the directory if it doesn't exist
if not os.path.exists(f"{self._save_folder}"):
os.makedirs(f"{self._save_folder}")
if self._index is None:
return
self._index.save_index(f"{self._save_folder}/index_{self._id}.bin")
# pickle the mappers
# Use old filenames for backwards compatibility
with open(f"{self._save_folder}/id_to_uuid_{self._id}.pkl", "wb") as f:
pickle.dump(self._label_to_id, f, pickle.HIGHEST_PROTOCOL)
with open(f"{self._save_folder}/uuid_to_id_{self._id}.pkl", "wb") as f:
pickle.dump(self._id_to_label, f, pickle.HIGHEST_PROTOCOL)
with open(f"{self._save_folder}/index_metadata_{self._id}.pkl", "wb") as f:
pickle.dump(self._index_metadata, f, pickle.HIGHEST_PROTOCOL)
logger.debug(f"Index saved to {self._save_folder}/index.bin")
def _exists(self) -> None:
return
def _load(self, curr_elements: int) -> None:
if not os.path.exists(f"{self._save_folder}/index_{self._id}.bin"):
return
# unpickle the mappers
with open(f"{self._save_folder}/id_to_uuid_{self._id}.pkl", "rb") as f:
self._label_to_id = pickle.load(f)
with open(f"{self._save_folder}/uuid_to_id_{self._id}.pkl", "rb") as f:
self._id_to_label = pickle.load(f)
with open(f"{self._save_folder}/index_metadata_{self._id}.pkl", "rb") as f:
self._index_metadata = pickle.load(f)
self._index_metadata["curr_elements"] = curr_elements
# Backwards compatability with versions that don't have curr_elements or total_elements_added
if "total_elements_added" not in self._index_metadata:
self._index_metadata["total_elements_added"] = self._index_metadata[
"elements"
]
p = hnswlib.Index(
space=self._params.space, dim=self._index_metadata["dimensionality"]
)
self._index = p
self._index.load_index(
f"{self._save_folder}/index_{self._id}.bin",
max_elements=int(
max(curr_elements * self._params.resize_factor, DEFAULT_CAPACITY)
),
)
self._index.set_ef(self._params.search_ef)
self._index.set_num_threads(self._params.num_threads)
def get_nearest_neighbors(
self, query: Embeddings, k: int, ids: Optional[List[UUID]] = None
) -> Tuple[List[List[UUID]], List[List[float]]]:
# The only case where the index is none is if no elements have been added
# We don't save the index until at least one element has been added
# And so there is also nothing at load time for persisted indexes
# In the case where no elements have been added, we return empty
if self._index is None:
return [[] for _ in range(len(query))], [[] for _ in range(len(query))]
# Check dimensionality
self._check_dimensionality(query)
# Check Number of requested results
if k > self._index_metadata["curr_elements"]:
logger.warning(
f"Number of requested results {k} is greater than number of elements in index {self._index_metadata['curr_elements']}, updating n_results = {self._index_metadata['curr_elements']}"
)
k = self._index_metadata["curr_elements"]
s2 = time.time()
# get ids from uuids as a set, if they are available
labels: Set[int] = set()
if ids is not None:
labels = {self._id_to_label[hexid(id)] for id in ids}
if len(labels) < k:
k = len(labels)
filter_function = None
if len(labels) != 0:
filter_function = lambda label: label in labels # NOQA: E731
logger.debug(f"time to pre process our knn query: {time.time() - s2}")
s3 = time.time()
database_labels, distances = self._index.knn_query(
query, k=k, filter=filter_function
)
distances = distances.tolist()
distances = cast(List[List[float]], distances)
logger.debug(f"time to run knn query: {time.time() - s3}")
return_ids = [
[self._label_to_id[label] for label in labels] for labels in database_labels
]
return return_ids, distances