forked from facebookresearch/fairseq-lua
-
Notifications
You must be signed in to change notification settings - Fork 0
/
webapi.lua
313 lines (284 loc) · 10.9 KB
/
webapi.lua
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
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
-- Copyright (c) 2017-present, Facebook, Inc.
-- All rights reserved.
--
-- This source code is licensed under the license found in the LICENSE file in
-- the root directory of this source tree. An additional grant of patent rights
-- can be found in the PATENTS file in the same directory.
--
--[[
--
-- Hypothesis generation script with text file input, processed line-by-line.
-- By default, this will run in interactive mode.
--
--]]
require 'fairseq'
local tnt = require 'torchnet'
local tds = require 'tds'
local argcheck = require 'argcheck'
local plstringx = require 'pl.stringx'
local data = require 'fairseq.torchnet.data'
local search = require 'fairseq.search'
local tokenizer = require 'fairseq.text.tokenizer'
local mutils = require 'fairseq.models.utils'
local cmd = torch.CmdLine()
cmd:option('-path', 'model1.th7,model2.th7', 'path to saved model(s)')
cmd:option('-beam', 1, 'search beam width')
cmd:option('-lenpen', 1,
'length penalty: <1.0 favors shorter, >1.0 favors longer sentences')
cmd:option('-unkpen', 0,
'unknown word penalty: <0 produces more, >0 produces less unknown words')
cmd:option('-subwordpen', 0,
'subword penalty: <0 favors longer, >0 favors shorter words')
cmd:option('-covpen', 0,
'coverage penalty: favor hypotheses that cover all source tokens')
cmd:option('-nbest', 1, 'number of candidate hypotheses')
cmd:option('-minlen', 1, 'minimum length of generated hypotheses')
cmd:option('-maxlen', 500, 'maximum length of generated hypotheses')
cmd:option('-input', '-', 'source language input text file')
cmd:option('-sourcedict', '', 'source language dictionary')
cmd:option('-targetdict', '', 'target language dictionary')
cmd:option('-vocab', '', 'restrict output to target vocab')
cmd:option('-visdom', '', 'visualize with visdom: (host:port)')
cmd:option('-model', '', 'model type for legacy models')
cmd:option('-aligndictpath', '', 'path to an alignment dictionary (optional)')
cmd:option('-nmostcommon', 500,
'the number of most common words to keep when using alignment')
cmd:option('-topnalign', 100, 'the number of the most common alignments to use')
cmd:option('-freqthreshold', -1,
'the minimum frequency for an alignment candidate in order' ..
'to be considered (default no limit)')
cmd:option('-fconvfast', false, 'make fconv model faster')
cmd:option('-unkaligndict', '', 'path to alignment dictionary')
cmd:option('-unkmarker', '<unk>', 'unknown word marker')
cmd:option('-offset', 0, 'apply offset to attention maxima')
local config = cmd:parse(arg)
-------------------------------------------------------------------
-- Load data
-------------------------------------------------------------------
config.dict = torch.load(config.targetdict)
print(string.format('| [target] Dictionary: %d types', config.dict:size()))
config.srcdict = torch.load(config.sourcedict)
print(string.format('| [source] Dictionary: %d types', config.srcdict:size()))
if config.aligndictpath ~= '' then
config.aligndict = tnt.IndexedDatasetReader{
indexfilename = config.aligndictpath .. '.idx',
datafilename = config.aligndictpath .. '.bin',
mmap = true,
mmapidx = true,
}
config.nmostcommon = math.max(config.nmostcommon, config.dict.nspecial)
config.nmostcommon = math.min(config.nmostcommon, config.dict:size())
end
local unkaligndict = torch.load(config.unkaligndict)
local TextFileIterator, _ =
torch.class('tnt.TextFileIterator', 'tnt.DatasetIterator', tnt)
local inputline
function print_r ( t )
local print_r_cache={}
local function sub_print_r(t,indent)
if (print_r_cache[tostring(t)]) then
print(indent.."*"..tostring(t))
else
print_r_cache[tostring(t)]=true
if (type(t)=="table") then
for pos,val in pairs(t) do
if (type(val)=="table") then
print(indent.."["..pos.."] => "..tostring(t).." {")
sub_print_r(val,indent..string.rep(" ",string.len(pos)+8))
print(indent..string.rep(" ",string.len(pos)+6).."}")
elseif (type(val)=="string") then
print(indent.."["..pos..'] => "'..val..'"')
else
print(indent.."["..pos.."] => "..tostring(val))
end
end
else
print(indent..tostring(t))
end
end
end
if (type(t)=="table") then
print(tostring(t).." {")
sub_print_r(t," ")
print("}")
else
sub_print_r(t," ")
end
print()
end
function transform1(line)
return {
bin = tokenizer.tensorizeString(line, config.srcdict),
text = line,
}
end
function transform2(sample)
local source = sample.bin:view(-1, 1):int()
local sourcePos = data.makePositions(source,
config.srcdict:getPadIndex()):view(-1, 1)
local sample = {
source = source,
sourcePos = sourcePos,
text = sample.text,
target = torch.IntTensor(1, 1), -- a stub
}
if config.aligndict then
sample.targetVocab, sample.targetVocabMap,
sample.targetVocabStats
= data.getTargetVocabFromAlignment{
dictsize = config.dict:size(),
unk = config.dict:getUnkIndex(),
aligndict = config.aligndict,
set = 'test',
source = sample.source,
target = sample.target,
nmostcommon = config.nmostcommon,
topnalign = config.topnalign,
freqthreshold = config.freqthreshold,
}
end
return sample
end
local model
if config.model ~= '' then
model = mutils.loadLegacyModel(config.path, config.model)
else
model = require(
'fairseq.models.ensemble_model'
).new(config)
if config.fconvfast then
local nfconv = 0
for _, fconv in ipairs(model.models) do
if torch.typename(fconv) == 'FConvModel' then
fconv:makeDecoderFast()
nfconv = nfconv + 1
end
end
assert(nfconv > 0, '-fconvfast requires an fconv model in the ensemble')
end
end
local vocab = nil
if config.vocab ~= '' then
vocab = tds.Hash()
local fd = io.open(config.vocab)
while true do
local line = fd:read()
if line == nil then
break
end
-- Add word on this line together with all prefixes
for i = 1, line:len() do
vocab[line:sub(1, i)] = 1
end
end
end
local searchf
if config.visdom ~= '' then
local host, port = table.unpack(plstringx.split(config.visdom, ':'))
searchf = search.visualize{
sf = searchf,
dict = config.dict,
sourceDict = config.srcdict,
host = host,
port = tonumber(port),
}
end
local dict, srcdict = config.dict, config.srcdict
local eos = dict:getSymbol(dict:getEosIndex())
local seos = srcdict:getSymbol(srcdict:getEosIndex())
local unk = dict:getSymbol(dict:getUnkIndex())
-- Select unknown token for reference that can't be produced by the model so
-- that the program output can be scored correctly.
local runk = unk
repeat
runk = string.format('<%s>', runk)
until dict:getIndex(runk) == dict:getUnkIndex()
function fix_sent(sample)
sample.bsz = 1
local hypos, scores, attns = model:generate(config, sample, searchf)
-- Print results
local hypos_unk_replace = {}
for i = 1, math.min(config.nbest, config.beam) do
local hypo = config.dict:getString(hypos[i]):gsub(eos .. '.*', '')
local htoks = plstringx.split(hypo)
local stoks = plstringx.split(inputline)
-- NOTE: This will print #hypo + 1 attention maxima. The last one is the
-- attention that was used to generate the <eos> symbol.
local _, maxattns = torch.max(attns[i], 2) -- attns (bsz * beam) * (targetlen * sourcelen)
local attens = maxattns:squeeze(2):totable()
for j = 1, #htoks do
if htoks[j] == config.unkmarker then
local attn = attens[j] + config.offset
if attn == #stoks + 1 then
if j == 1 then
htoks[j] = stoks[1]
else
htoks[j] = ''
end
elseif attn < 1 or attn > #stoks + 1 then
io.stderr:write(string.format(
'Sentence %d: attention index out of bound: %d\n',
i, attn))
else
local stok = stoks[attn]
if unkaligndict[stok] then
htoks[j] = unkaligndict[stok]
else
htoks[j] = stok
end
end
end
end
-- local res = {}
-- table.insert(res, plstringx.join(' ', htoks))
-- scores_cut = string.format("%.6f", scores[i])
-- table.insert(res, scores_cut)
-- table.insert(hypos_unk_replace, res)
-- print('H:', plstringx.join(' ', htoks))
table.insert(hypos_unk_replace, plstringx.join(' ', htoks))
print('H:', plstringx.join(' ', htoks))
end
return hypos_unk_replace
end
local app = require('waffle')
local base64 = require('base64')
local json = require('cjson')
function decodeURI(s)
s = string.gsub(s, '%%(%x%x)', function(h) return string.char(tonumber(h, 16)) end)
return s
end
app.get('/fairseq', function(req, res)
print_r(req)
local num = tonumber(req.url.args.num)
-- local str = string.sub(req.url.path, 5)
-- str = base64.decode(str)
-- local separate = string.find(str, '+')
-- local num = tonumber(string.sub(str, 0, separate - 1))
config.beam = num
config.nbest = num
searchf = search.beam{
ttype = model:type(),
dict = config.dict,
srcdict = config.srcdict,
beam = config.beam,
lenPenalty = config.lenpen,
unkPenalty = config.unkpen,
subwordPenalty = config.subwordpen,
coveragePenalty = config.covpen,
vocab = vocab,
}
-- local question = string.sub(str, separate + 1)
local question = req.url.args.q
-- question = decodeURI(question)
question = string.gsub(question, '+', ' ')
print(question)
inputline = question
question = transform1(question)
question = transform2(question)
question = fix_sent(question)
question = json.encode(question)
res.header('Access-Control-Allow-Origin', '*')
res.header('Content-Type', 'application/json')
res.json(question)
end)
app.listen({host='10.190.190.137', port='8080'})