File size: 9,163 Bytes
05d3571
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/python3
# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
#
# LASER  Language-Agnostic SEntence Representations
# is a toolkit to calculate multilingual sentence embeddings
# and to use them for document classification, bitext filtering
# and mining
#
# --------------------------------------------------------
#
# tools for indexing and search with FAISS

import faiss
import os.path
import sys
import numpy as np

#-------------------------------------------------------------
# Get list of fnames:
#  - we loop over the list of given languages
#  - for each language, we also check if there are splitted files .%03d

def SplitFnames(par_fname, langs):
    fnames = []
    for l in langs:
        fname = par_fname + '.' + l
        if os.path.isfile(fname):
            fnames.append(fname)
        for i in range(1000):
            fname = par_fname + '.' + l + '.{:03d}'.format(i)
            if os.path.isfile(fname):
                fnames.append(fname)
    if len(fnames) == 0:
        print("ERROR: no embeddings found in {:s}*".format(par_fname))
        sys.exit(1)
    return fnames

def SplitOpen(par_fname, langs, dim, dtype, verbose=False):
    M = []
    nf = 0
    nc = 0
    print('Reading sentence embeddings')
    print(' - memory mapped files {:s}'.format(par_fname))
    for fname in SplitFnames(par_fname, langs):
        n = int(os.path.getsize(fname) / dim / np.dtype(dtype).itemsize)
        if verbose:
            print(' - {:s}: {:d} x {:d}'.format(fname, n, dim))
        Mi = np.memmap(fname, mode='r', dtype=dtype, shape=(n, dim))
        nc += n
        nf += 1
        M.append(Mi)
    print(' - total of {:d} files: {:d} x {:d}'.format(nf, nc, dim))
    return M

def SplitAccess(M, idx):
    i = idx
    for Mi in M:
        n = Mi.shape[0]
        if i < n:
            return Mi[i,:]
        i -= n
    print('ERROR: index {:d} is too large form memory mapped files'.format(idx))
    sys.exit(1)


###############################################################################
# create an FAISS index on the given data

def IndexCreate(dname, idx_type,
                verbose=False, normalize=True, save_index=False, dim=1024):

    assert idx_type == 'FlatL2', 'only FlatL2 index is currently supported'
    x = np.fromfile(dname, dtype=np.float32, count=-1)
    nbex = x.shape[0] // dim
    print(' - embedding: {:s} {:d} examples of dim {:d}'
          .format(dname, nbex, dim))
    x.resize(nbex, dim)
    print(' - creating FAISS index')
    idx = faiss.IndexFlatL2(dim)
    if normalize:
        faiss.normalize_L2(x)
    idx.add(x)
    if save_index:
        iname = 'TODO'
        print(' - saving index into ' + iname)
        faiss.write_index(idx, iname)
    return x, idx


###############################################################################
# search closest vector for all languages pairs and calculate error rate

def IndexSearchMultiple(data, idx, langs, verbose=False, texts=None, print_errors=False):
    nl = len(data)
    nbex = data[0].shape[0]
    err = np.zeros((nl, nl)).astype(float)
    ref = np.linspace(0, nbex-1, nbex).astype(int)  # [0, nbex)
    if verbose:
        if texts is None:
            print('Calculating similarity error (indices):')
        else:
            print('Calculating similarity error (textual):')
    for i1 in range(nl):
        for i2 in range(nl):
            if i1 != i2:
                D, I = idx[i2].search(data[i1], 1)
                if texts: # do textual comparison
                    e1 = 0
                    for p in range(I.shape[0]):
                        if texts[i2][p] != texts[i2][I[p,0]]:
                            e1 += 1
                            if print_errors:
                                print('Error {:s}\n      {:s}'
                                      .format(texts[i2][p].strip(), texts[i2][I[p,0]].strip()))
                    err[i1, i2] = e1 / nbex
                else:  # do index based comparision
                    err[i1, i2] \
                        = (nbex - np.equal(I.reshape(nbex), ref)
                           .astype(int).sum()) / nbex
                if verbose:
                    print(' - similarity error {:s}/{:s}: {:5.2f}%'
                          .format(langs[i1], langs[i2],
                                  100.0 * err[i1, i2]))
    return err


###############################################################################
# print confusion matrix

def IndexPrintConfusionMatrix(err, langs):
    nl = len(langs)
    assert nl == err.shape[0], 'size of errror matrix doesn not match'
    print('Confusion matrix:')
    print('{:8s}'.format('langs'), end='')
    for i2 in range(nl):
        print('{:8s} '.format(langs[i2]), end='')
    print('{:8s}'.format('avg'))
    for i1 in range(nl):
        print('{:3s}'.format(langs[i1]), end='')
        for i2 in range(nl):
            print('{:8.2f}%'.format(100 * err[i1, i2]), end='')
        print('{:8.2f}%'.format(100 * err[i1, :].sum() / (nl-1)))

    print('avg', end='')
    for i2 in range(nl):
        print('{:8.2f}%'.format(100 * err[:, i2].sum() / (nl-1)), end='')

    # global average
    print('{:8.2f}%'.format(100 * err.sum() / (nl-1) / nl))


###############################################################################
# Load an FAISS index

def IndexLoad(idx_name, nprobe, gpu=False):
    print('Reading FAISS index')
    print(' - index: {:s}'.format(idx_name))
    index = faiss.read_index(idx_name)
    print(' - found {:d} sentences of dim {:d}'.format(index.ntotal, index.d))
    print(' - setting nbprobe to {:d}'.format(nprobe))
    if gpu:
        print(' - transfer index to %d GPUs ' % faiss.get_num_gpus())
        #co = faiss.GpuMultipleClonerOptions()
        #co.shard = True
        index = faiss.index_cpu_to_all_gpus(index) # co=co
        faiss.GpuParameterSpace().set_index_parameter(index, 'nprobe', nprobe)
    return index


###############################################################################
# Opens a text file with the sentences corresponding to the indices used
# by an FAISS index
# We also need the reference files with the byte offsets to the beginning
# of each sentence
# optionnally:  array with number of words per sentence
# All arrays are memory mapped

def IndexTextOpen(txt_fname):
    print('Reading text corpus')
    print(' - texts: {:s}'.format(txt_fname))
    txt_mmap = np.memmap(txt_fname, mode='r', dtype=np.uint8)
    fname = txt_fname.replace('.txt', '.ref.bin32')
    if os.path.isfile(fname):
        print(' - sentence start offsets (32 bit): {}'.format(fname))
        ref_mmap = np.memmap(fname, mode='r', dtype=np.uint32)
    else:
        fname = txt_fname.replace('.txt', '.ref.bin64')
        if os.path.isfile(fname):
            print(' - sentence start offsets (64 bit): {}'.format(fname))
            ref_mmap = np.memmap(fname, mode='r', dtype=np.uint64)
        else:
            print('ERROR: no file with sentence start offsets found')
            sys.exit(1)
    print(' - found {:d} sentences'.format(ref_mmap.shape[0]))

    nbw_mmap = None
    fname = txt_fname.replace('.txt', '.nw.bin8')
    if os.path.isfile(fname):
        print(' - word counts: {:s}'.format(fname))
        nbw_mmap = np.memmap(fname, mode='r', dtype=np.uint8)

    M = None
    fname = txt_fname.replace('.txt', '.meta')
    if os.path.isfile(fname):
        M = []
        n = 0
        print(' - metafile: {:s}'.format(fname))
        with open(fname, 'r') as fp:
            for line in fp:
                fields = line.strip().split()
                if len(fields) != 2:
                    print('ERROR: format error in meta file')
                    sys.exit(1)
                n += int(fields[1])
                M.append({'lang': fields[0], 'n': n})
        print(' - found {:d} languages:'.format(len(M)), end='')
        for L in M:
            print(' {:s}'.format(L['lang']), end='')
        print('')

    return txt_mmap, ref_mmap, nbw_mmap, M


###############################################################################
# Return the text for the given index

def IndexTextQuery(txt_mmap, ref_mmap, idx):
    p = int(ref_mmap[idx])  # get starting byte position
    i = 0
    dim = 10000  # max sentence length in bytes
    b = bytearray(dim)
    #  find EOL
    while txt_mmap[p+i] != 10 and i < dim:
        b[i] = txt_mmap[p+i]
        i += 1

    return b[0:i].decode('utf-8')


###############################################################################
# Search the [k] nearest vectors of [x] in the given index
# and return the text lines

def IndexSearchKNN(index, x, T, R, kmax=1, Dmax=1.0, dedup=True):
    D, I = index.search(x, kmax)
    prev = {}  # for depuplication
    res = []
    for n in range(x.shape[0]):
        for i in range(kmax):
            txt = IndexTextQuery(T, R, I[n, i])
            if (dedup and txt not in prev) and D[n, i] <= Dmax:
                prev[txt] = 1
                res.append([txt, D[n, i]])
    return res