-
Notifications
You must be signed in to change notification settings - Fork 7
/
train.lua
199 lines (155 loc) · 5.71 KB
/
train.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
require('onmt.init')
require('tds')
local cmd = onmt.utils.ExtendedCmdLine.new('train.lua')
-- First argument define the model type: seq2seq/lm - default is seq2seq.
local modelType = cmd.getArgument(arg, '-model_type') or 'seq2seq'
local modelClass = onmt.ModelSelector(modelType)
-- Options declaration.
local options = {
{
'-data', '',
[[Path to the data package `*-train.t7` generated by the preprocessing step.]],
{
valid = onmt.utils.ExtendedCmdLine.fileNullOrExists
}
}
}
cmd:setCmdLineOptions(options, 'Data')
onmt.data.SampledDataset.declareOpts(cmd)
onmt.data.DynamicDataRepository.declareOpts(cmd, modelClass)
onmt.data.SampledVocabDataset.declareOpts(cmd)
onmt.Model.declareOpts(cmd)
modelClass.declareOpts(cmd)
onmt.train.Trainer.declareOpts(cmd)
onmt.utils.CrayonLogger.declareOpts(cmd)
onmt.utils.Cuda.declareOpts(cmd)
onmt.utils.Logger.declareOpts(cmd)
cmd:text('')
cmd:text('Other options')
cmd:text('')
onmt.utils.Memory.declareOpts(cmd)
onmt.utils.Profiler.declareOpts(cmd)
cmd:option('-seed', 3435, [[Random seed.]], {valid=onmt.utils.ExtendedCmdLine.isUInt()})
local function loadData(opt, filename)
local data
if filename ~= '' then
_G.logger:info('Loading data from \'%s\'...', filename)
data = torch.load(filename, 'binary', false)
-- Check if data type is compatible with the target model.
onmt.utils.Error.assert(modelClass.dataType(data.dataType),
'Data type `%s\' is incompatible with `%s\' models',
data.dataType, modelClass.modelName())
else
data = onmt.data.DynamicDataRepository.new(opt, modelClass)
end
-- Keep backward compatibility.
data.dataType = data.dataType or 'bitext'
return data
end
local function buildDataset(opt, data)
local trainDataset, validDataset
if torch.type(data) == "DynamicDataRepository" then
validDataset = data:getValid()
trainDataset = data:getTraining()
else
if opt.sample > 0 then
trainDataset = onmt.data.SampledDataset.new(opt, data.train.src, data.train.tgt)
else
trainDataset = onmt.data.Dataset.new(data.train.src, data.train.tgt)
end
if data.valid then
validDataset = onmt.data.Dataset.new(data.valid.src, data.valid.tgt)
end
end
local nTrainBatch, batchUsage = trainDataset:setBatchSize(opt.max_batch_size, opt.uneven_batches)
if validDataset then
validDataset:setBatchSize(opt.max_batch_size, opt.uneven_batches)
else
_G.logger:warning('No validation data')
end
if data.dataType ~= 'monotext' then
local srcVocSize
local srcFeatSize = '-'
if data.dicts.src then
srcVocSize = data.dicts.src.words:size()
srcFeatSize = #data.dicts.src.features
else
srcVocSize = '*'..data.dicts.srcInputSize
end
local tgtVocSize
local tgtFeatSize = '-'
if data.dicts.tgt then
tgtVocSize = data.dicts.tgt.words:size()
tgtFeatSize = #data.dicts.tgt.features
else
tgtVocSize = '*'..data.dicts.tgtInputSize
end
_G.logger:info(' * vocabulary size: source = %s; target = %s',
srcVocSize, tgtVocSize)
_G.logger:info(' * additional features: source = %s; target = %s',
srcFeatSize, tgtFeatSize)
else
_G.logger:info(' * vocabulary size: %d', data.dicts.src.words:size())
_G.logger:info(' * additional features: %d', #data.dicts.src.features)
end
_G.logger:info(' * maximum sequence length: source = %d; target = %d',
trainDataset.maxSourceLength, trainDataset.maxTargetLength)
_G.logger:info(' * number of training sentences: %d', #trainDataset.src)
_G.logger:info(' * number of batches: %d', nTrainBatch)
_G.logger:info(' - source sequence lengths: %s', opt.uneven_batches and 'variable' or 'equal')
_G.logger:info(' - maximum size: %d', opt.max_batch_size)
_G.logger:info(' - average size: %.2f', trainDataset:instanceCount() / nTrainBatch)
_G.logger:info(' - capacity: %.2f%%', math.ceil(batchUsage * 1000) / 10)
return trainDataset, validDataset
end
local function loadModel(opt, dicts)
local checkpoint
local paramChanges
checkpoint, opt, paramChanges = onmt.train.Saver.loadCheckpoint(opt)
cmd:logConfig(opt)
local model = modelClass.load(opt, checkpoint.models, dicts)
-- Change parameters dynamically.
if not onmt.utils.Table.empty(paramChanges) then
model:changeParameters(paramChanges)
end
return model, checkpoint.info
end
local function buildModel(opt, dicts)
_G.logger:info('Building model...')
return modelClass.new(opt, dicts)
end
local function main()
local opt = cmd:parse(arg)
torch.manualSeed(opt.seed)
-- Initialize global context.
_G.logger = onmt.utils.Logger.new(opt.log_file, opt.disable_logs, opt.log_level)
_G.crayon_logger = onmt.utils.CrayonLogger.new(opt)
_G.profiler = onmt.utils.Profiler.new(false)
onmt.utils.Cuda.init(opt)
onmt.utils.Parallel.init(opt)
_G.logger:info('Training ' .. modelClass.modelName() .. ' model...')
-- Loading data package.
local data = loadData(opt, opt.data)
-- Record data type in the options, and preprocessing options if present.
opt.data_type = data.dataType
opt.preprocess = data.opt
-- Building training datasets.
local trainDataset, validDataset = buildDataset(opt, data)
-- Building the model.
local model
local trainStates
if onmt.train.Saver.checkpointDefined(opt) then
model, trainStates = loadModel(opt, data.dicts)
else
model = buildModel(opt, data.dicts)
end
onmt.utils.Cuda.convert(model)
if opt.sample > 0 then
trainDataset:checkModel(model)
end
-- Start training.
local trainer = onmt.train.Trainer.new(opt, model, data.dicts, trainDataset:getBatch(1))
trainer:train(trainDataset, validDataset, trainStates)
_G.logger:shutDown()
end
main()