milvus/pyengine/engine/controller/scheduler.py

72 lines
2.4 KiB
Python

from engine.retrieval import search_index
from engine.ingestion import build_index
from engine.ingestion import serialize
import numpy as np
class Singleton(type):
_instances = {}
def __call__(cls, *args, **kwargs):
if cls not in cls._instances:
cls._instances[cls] = super(Singleton, cls).__call__(*args, **kwargs)
return cls._instances[cls]
class Scheduler(metaclass=Singleton):
def search(self, index_file_key, vectors, k):
# assert index_file_key
# assert vectors
assert k != 0
query_vectors = serialize.to_array(vectors)
return self.__scheduler(index_file_key, query_vectors, k)
def __scheduler(self, index_data_key, vectors, k):
result_list = []
if 'raw' in index_data_key:
raw_vectors = index_data_key['raw']
raw_vector_ids = index_data_key['raw_id']
d = index_data_key['dimension']
index_builder = build_index.FactoryIndex()
index = index_builder().build(d, raw_vectors, raw_vector_ids)
searcher = search_index.FaissSearch(index)
result_list.append(searcher.search_by_vectors(vectors, k))
if 'index' in index_data_key:
index_data_list = index_data_key['index']
for key in index_data_list:
index = get_index_data(key)
searcher = search_index.FaissSearch(index)
result_list.append(searcher.search_by_vectors(vectors, k))
if len(result_list) == 1:
return result_list[0].vectors[0].tolist() # TODO(linxj): fix hard code
return result_list; # TODO(linxj): add topk
# d_list = np.array([])
# v_list = np.array([])
# for result in result_list:
# rd = result.distance
# rv = result.vectors
#
# td_list = np.array([])
# tv_list = np.array([])
# for d, v in zip(rd, rv):
# td_list = np.append(td_list, d)
# tv_list = np.append(tv_list, v)
# d_list = np.add(d_list, td_list)
# v_list = np.add(v_list, td_list)
#
# print(d_list)
# print(v_list)
# result_map = [d_list, v_list]
# top_k_result = search_index.top_k(result_map, k)
# return top_k_result
def get_index_data(key):
return serialize.read_index(key)