-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_datastore.py
134 lines (105 loc) · 5.18 KB
/
train_datastore.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
import argparse
import os
import numpy as np
import faiss
import time
import ctypes
overall_start = time.time()
parser = argparse.ArgumentParser()
parser.add_argument('--dstore_mmap', type=str, help='memmap where keys and vals are stored')
parser.add_argument('--dstore_size', type=int, help='number of items saved in the datastore memmap')
parser.add_argument('--chunk_size', type=int, default=1, help='number of items saved in the datastore memmap')
parser.add_argument('--dimension', type=int, default=1024, help='Size of each key')
parser.add_argument('--dstore-fp16', default=False, action='store_true')
parser.add_argument('--seed', type=int, default=1,
help='random seed for sampling the subset of vectors to train the cache')
parser.add_argument('--ncentroids', type=int, default=4096, help='number of centroids faiss should learn')
parser.add_argument('--code_size', type=int, default=64, help='size of quantized vectors')
parser.add_argument('--probe', type=int, default=32, help='number of clusters to query')
parser.add_argument('--faiss_index', type=str, help='file to write the faiss index')
parser.add_argument('--num_keys_to_add_at_a_time', default=1000000, type=int,
help='can only load a certain amount of data to memory at a time.')
parser.add_argument('--starting_point', type=int, default=0, help='index to start adding keys at')
parser.add_argument('--use_gpu', default=False, action='store_true')
parser.add_argument("--pca", default=0, type=int)
args = parser.parse_args()
print(args)
res = faiss.StandardGpuResources()
# load the saved keys and values
if args.dstore_fp16:
print('load dstore fp16', args.dstore_size, args.dimension)
keys = np.memmap(args.dstore_mmap + 'keys.npy', dtype=np.float16, mode='r', shape=(args.dstore_size, args.dimension))
vals = np.memmap(args.dstore_mmap + 'vals.npy', dtype=np.int, mode='r', shape=(args.dstore_size, args.chunk_size))
else:
print(args.dstore_mmap + 'keys.npy')
keys = np.memmap(args.dstore_mmap + 'keys.npy', dtype=np.float32, mode='r', shape=(args.dstore_size, args.dimension))
vals = np.memmap(args.dstore_mmap + 'vals.npy', dtype=np.int, mode='r', shape=(args.dstore_size, args.chunk_size))
print('done.')
# to speed up access to np.memmap
madvise = ctypes.CDLL("libc.so.6").madvise
madvise.argtypes = [ctypes.c_void_p, ctypes.c_size_t, ctypes.c_int]
madvise.restype = ctypes.c_int
assert madvise(keys.ctypes.data, keys.size * keys.dtype.itemsize, 1) == 0, "MADVISE FAILED" # 2 means MADV_SEQUENTIAL
if not os.path.exists(args.faiss_index + ".trained"):
index_dim = args.pca if args.pca > 0 else args.dimension
# Initialize faiss index
quantizer = faiss.IndexFlatL2(index_dim)
index = faiss.IndexIVFPQ(quantizer, index_dim, args.ncentroids, args.code_size, 8)
index.nprobe = args.probe
if args.use_gpu:
print('Start put index to gpu')
co = faiss.GpuClonerOptions()
co.useFloat16 = True
gpu_index = faiss.index_cpu_to_gpu(res, 0, index, co)
if args.pca > 0:
pca_matrix = faiss.PCAMatrix(args.dimension, args.pca, 0, True)
index = faiss.IndexPreTransform(pca_matrix, index)
print('Training Index')
np.random.seed(args.seed)
random_sample = np.random.choice(np.arange(vals.shape[0]), size=[min(1000000, vals.shape[0])],replace=False)
start = time.time()
if args.use_gpu:
gpu_index.train(keys[random_sample].astype(np.float32))
else:
index.train(keys[random_sample].astype(np.float32))
print('Training took {} s'.format(time.time() - start))
print('Writing index after training')
start = time.time()
if args.use_gpu:
faiss.write_index(faiss.index_gpu_to_cpu(gpu_index), args.faiss_index + ".trained")
else:
faiss.write_index(index, args.faiss_index+".trained")
print('Writing index took {} s'.format(time.time() - start))
print('Adding Keys')
index = faiss.read_index(args.faiss_index + ".trained")
if args.use_gpu:
co = faiss.GpuClonerOptions()
co.useFloat16 = True
gpu_index = faiss.index_cpu_to_gpu(res, 0, index, co)
start = args.starting_point
start_time = time.time()
while start < args.dstore_size:
end = min(args.dstore_size, start + args.num_keys_to_add_at_a_time)
to_add = keys[start:end].copy()
if args.use_gpu:
gpu_index.add_with_ids(to_add.astype(np.float32), np.arange(start, end))
else:
index.add_with_ids(to_add.astype(np.float32), np.arange(start, end))
start += args.num_keys_to_add_at_a_time
if (start % 1000000) == 0:
print('Added %d tokens so far' % start)
print('Writing Index', start)
if args.use_gpu:
faiss.write_index(faiss.index_gpu_to_cpu(gpu_index), args.faiss_index)
else:
faiss.write_index(index, args.faiss_index)
print("Adding total %d keys" % end)
print('Adding took {} s'.format(time.time() - start_time))
print('Writing Index')
start = time.time()
if args.use_gpu:
faiss.write_index(faiss.index_gpu_to_cpu(gpu_index), args.faiss_index)
else:
faiss.write_index(index, args.faiss_index)
print('Writing index took {} s'.format(time.time() - start))
print('Training datastore took {} s'.format(time.time() - overall_start))