forked from NNDam/faiss
-
Notifications
You must be signed in to change notification settings - Fork 0
/
6-Test-faiss-quality.py
289 lines (204 loc) · 9.08 KB
/
6-Test-faiss-quality.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
import os
import faiss
import numpy as np
from typing import List
np.random.seed(42)
d = 512
nb = 4096
xb = np.random.rand(nb, d).astype(np.float32)
r_qua = np.random.rand(nb, 1).astype(np.float32)
ids = np.array(range(nb))
nlist = 16
k = 5
folder_save_onram_wo_qua = "./testing/saved/onram/wo_qua"
folder_save_onram_w_qua = "./testing/saved/onram/w_qua"
folder_save_ondisk = "./testing/saved/ondisk/w_qua"
def merge_ondisk(
trained_index: faiss.Index, shard_fnames: List[str], ivfdata_fname: str
) -> None:
"""Add the contents of the indexes stored in shard_fnames into the index
trained_index. The on-disk data is stored in ivfdata_fname"""
assert not isinstance(
trained_index, faiss.IndexIVFPQR
), "IndexIVFPQR is not supported as an on disk index."
# merge the images into an on-disk index
# first load the inverted lists
ivfs = []
for fname in shard_fnames:
# the IO_FLAG_MMAP is to avoid actually loading the data thus
# the total size of the inverted lists can exceed the
# available RAM
index = faiss.read_index(fname, faiss.IO_FLAG_MMAP)
index_ivf = faiss.extract_index_ivf(index)
ivfs.append(index_ivf.invlists)
# avoid that the invlists get deallocated with the index
index_ivf.own_invlists = False
# construct the output index
index = trained_index
index_ivf = faiss.extract_index_ivf(index)
assert index.ntotal == 0, "works only on empty index"
# prepare the output inverted lists. They will be written
# to merged_index.ivfdata
invlists = faiss.OnDiskInvertedLists(
index_ivf.nlist, index_ivf.code_size, ivfdata_fname
)
# merge all the inverted lists
ivf_vector = faiss.InvertedListsPtrVector()
for ivf in ivfs:
ivf_vector.push_back(ivf)
ntotal = invlists.merge_from(ivf_vector.data(), ivf_vector.size())
# now replace the inverted lists in the output index
index.ntotal = index_ivf.ntotal = ntotal
index_ivf.replace_invlists(invlists, True)
invlists.this.disown()
def compare_search(index_wo_qua, index_w_q, qua_1 = False, qua_2 = True):
if qua_1:
D1, I1, Q2 = index_wo_qua.search_with_quality(xb[:10], k, 0, 1.0)
else:
D1, I1 = index_wo_qua.search(xb[:10], k)
if qua_2:
D2, I2, Q2 = index_w_q.search_with_quality(xb[:10], k, 0, 1.0)
else:
D2, I2 = index_w_q.search(xb[:10], k)
assert np.sum(I1 - I2) == 0.0
assert np.sum(D1 - D2) == 0.0
def test_index_flat():
# Train
index_flat_wo_qua = faiss.IndexFlatL2(d)
index_flat_w_qua = faiss.IndexFlatL2(d)
index_flat_w_qua.set_include_quality()
# Search
index_flat_w_qua.add_with_quality(xb, r_qua)
index_flat_wo_qua.add(xb)
compare_search(index_flat_wo_qua, index_flat_w_qua)
# Save
name_save= "index_flat.bin"
os.makedirs(folder_save_onram_wo_qua, exist_ok = True)
os.makedirs(folder_save_onram_w_qua, exist_ok = True)
faiss.write_index(index_flat_wo_qua, os.path.join(folder_save_onram_wo_qua, name_save))
faiss.write_index(index_flat_w_qua, os.path.join(folder_save_onram_w_qua, name_save))
backup_index_flat_wo_qua = index_flat_wo_qua
del index_flat_w_qua
del index_flat_wo_qua
index_flat_w_qua = faiss.read_index(os.path.join(folder_save_onram_w_qua, name_save))
index_flat_wo_qua = faiss.read_index(os.path.join(folder_save_onram_wo_qua, name_save))
compare_search(backup_index_flat_wo_qua, index_flat_w_qua)
compare_search(index_flat_wo_qua, index_flat_w_qua)
def test_index_ivfflat():
# Train
index_flat_wo_qua = faiss.index_factory(512, "IVF16,Flat")
index_flat_w_qua = faiss.index_factory(512, "IVF16,Flat")
index_flat_w_qua.set_include_quality()
# Search
if not index_flat_w_qua.is_trained:
index_flat_w_qua.train(xb)
if not index_flat_wo_qua.is_trained:
index_flat_wo_qua.train(xb)
index_flat_w_qua.add_with_quality(xb, r_qua)
index_flat_wo_qua.add(xb)
compare_search(index_flat_wo_qua, index_flat_w_qua)
# Save
name_save= "index_flat.bin"
os.makedirs(folder_save_onram_wo_qua, exist_ok = True)
os.makedirs(folder_save_onram_w_qua, exist_ok = True)
faiss.write_index(index_flat_wo_qua, os.path.join(folder_save_onram_wo_qua, name_save))
faiss.write_index(index_flat_w_qua, os.path.join(folder_save_onram_w_qua, name_save))
backup_index_flat_wo_qua = index_flat_wo_qua
del index_flat_w_qua
del index_flat_wo_qua
index_flat_w_qua = faiss.read_index(os.path.join(folder_save_onram_w_qua, name_save))
index_flat_wo_qua = faiss.read_index(os.path.join(folder_save_onram_wo_qua, name_save))
compare_search(backup_index_flat_wo_qua, index_flat_w_qua)
compare_search(index_flat_wo_qua, index_flat_w_qua)
def test_index_ivfflat_ondisk_w_qua():
# Train
index_flat_w_qua = faiss.index_factory(d, "IVF16,SQfp16")
index_flat_w_qua.set_include_quality()
# index_flat_w_qua = faiss.IndexIDMap(index_flat_w_qua)
index_flat_w_qua.train(xb)
index_ivf_part = faiss.extract_index_ivf(index_flat_w_qua)
clustering_index = faiss.IndexFlatL2(d)
clustering_index.reset()
index_ivf_part.clustering_index = clustering_index
if index_flat_w_qua.is_trained is False:
index_flat_w_qua.train(xb)
os.makedirs(folder_save_ondisk, exist_ok = True)
os.makedirs(folder_save_onram_w_qua, exist_ok = True)
name = "index_ivfsq"
faiss.write_index(index_flat_w_qua, os.path.join(folder_save_ondisk, name + ".bin"))
del index_flat_w_qua
index_flat_w_qua = faiss.read_index(os.path.join(folder_save_ondisk, name + ".bin"))
index_flat_w_qua.add_with_ids_with_quality(xb, r_qua, ids)
index_flat_w_qua.search_with_quality(xb[:10], 5, 0, 1.0)
faiss.write_index(index_flat_w_qua, os.path.join(folder_save_onram_w_qua, name + ".bin"))
del index_flat_w_qua
index_flat_w_qua_only_trained = faiss.read_index(os.path.join(folder_save_ondisk, name + ".bin"))
# from faiss.contrib.ondisk import merge_ondisk
merge_ondisk(
index_flat_w_qua_only_trained,
[os.path.join(folder_save_onram_w_qua, name + ".bin")],
os.path.join(folder_save_ondisk, name + ".ivfdata")
)
# exit()
print("[INFO] Write on disk index: ")
faiss.write_index(index_flat_w_qua_only_trained, os.path.join(folder_save_ondisk, name + ".bin"))
compare_search(
faiss.read_index(os.path.join(folder_save_onram_w_qua, name + ".bin")),
faiss.read_index(os.path.join(folder_save_ondisk, name + ".bin")),
True,
True
)
def test_index_idmap():
# Train
index_flat_wo_qua = faiss.index_factory(512, "IDMap,SQfp16")
index_flat_w_qua = faiss.index_factory(512, "IDMap,SQfp16")
index_flat_w_qua.set_include_quality()
# Search
if not index_flat_w_qua.is_trained:
index_flat_w_qua.train(xb)
if not index_flat_wo_qua.is_trained:
index_flat_wo_qua.train(xb)
index_flat_w_qua.add_with_ids_with_quality(xb, r_qua, ids)
index_flat_wo_qua.add_with_ids(xb, ids)
compare_search(index_flat_wo_qua, index_flat_w_qua)
# Save
name_save= "index_flat.bin"
os.makedirs(folder_save_onram_wo_qua, exist_ok = True)
os.makedirs(folder_save_onram_w_qua, exist_ok = True)
faiss.write_index(index_flat_wo_qua, os.path.join(folder_save_onram_wo_qua, name_save))
faiss.write_index(index_flat_w_qua, os.path.join(folder_save_onram_w_qua, name_save))
backup_index_flat_wo_qua = index_flat_wo_qua
del index_flat_w_qua
del index_flat_wo_qua
index_flat_w_qua = faiss.read_index(os.path.join(folder_save_onram_w_qua, name_save))
index_flat_wo_qua = faiss.read_index(os.path.join(folder_save_onram_wo_qua, name_save))
compare_search(backup_index_flat_wo_qua, index_flat_w_qua)
compare_search(index_flat_wo_qua, index_flat_w_qua)
def test_reconstruct_quality():
name_factory = "IVF16,Flat"
index_flat_w_qua = faiss.index_factory(512, name_factory)
index_flat_w_qua.set_include_quality()
index_flat_w_qua.make_direct_map()
index_flat_w_qua.set_direct_map_type(faiss.DirectMap.Hashtable)
# print(name_factory)
index_flat_w_qua.train(xb)
index_flat_w_qua.add_with_ids_with_quality(xb, r_qua, ids)
# index_flat_w_qua.add_with_quality(xb, r_qua)
D, I, Q = index_flat_w_qua.search_with_quality(xb[:10], k, 0, 1.0)
for idx_r, row in enumerate(I):
for idx_c, element in enumerate(row):
r_recons = index_flat_w_qua.reconstruct_qua(int(ids[element]))
true_r_qua = r_qua[ids[element]]
print(r_recons, true_r_qua, Q[idx_r][idx_c])
print(name_factory)
if __name__ == "__main__":
print("[INFO] Testing reconstruct quality ")
test_reconstruct_quality()
print("[INFO] Testing index flat L2: ")
test_index_flat()
print("[INFO] Testing index IVFFlatL2: ")
test_index_ivfflat()
print("[INFO] Testing index IVFFSQ ondisk: ")
test_index_ivfflat_ondisk_w_qua()
print("[INFO] Testing index IDMap: ")
test_index_idmap()