nnet-train-parallel.cc
3.73 KB
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
// nnet2bin/nnet-train-parallel.cc
// Copyright 2012 Johns Hopkins University (author: Daniel Povey)
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "hmm/transition-model.h"
#include "nnet2/nnet-update-parallel.h"
#include "nnet2/am-nnet.h"
int main(int argc, char *argv[]) {
try {
using namespace kaldi;
using namespace kaldi::nnet2;
typedef kaldi::int32 int32;
typedef kaldi::int64 int64;
const char *usage =
"Train the neural network parameters with backprop and stochastic\n"
"gradient descent using minibatches. As nnet-train-simple, but\n"
"uses multiple threads in a Hogwild type of update (for CPU, not GPU).\n"
"\n"
"Usage: nnet-train-parallel [options] <model-in> <training-examples-in> <model-out>\n"
"\n"
"e.g.:\n"
"nnet-train-parallel --num-threads=8 1.nnet ark:1.1.egs 2.nnet\n";
bool binary_write = true;
bool zero_stats = true;
int32 minibatch_size = 1024;
int32 srand_seed = 0;
ParseOptions po(usage);
po.Register("binary", &binary_write, "Write output in binary mode");
po.Register("zero-stats", &zero_stats, "If true, zero stats "
"stored with the neural net (only affects mixing up).");
po.Register("srand", &srand_seed,
"Seed for random number generator (e.g., for dropout)");
po.Register("num-threads", &g_num_threads, "Number of training threads to use "
"in the parallel update. [Note: if you use a parallel "
"implementation of BLAS, the actual number of threads may be larger.]");
po.Register("minibatch-size", &minibatch_size, "Number of examples to use for "
"each minibatch during training.");
po.Read(argc, argv);
srand(srand_seed);
if (po.NumArgs() != 3) {
po.PrintUsage();
exit(1);
}
std::string nnet_rxfilename = po.GetArg(1),
examples_rspecifier = po.GetArg(2),
nnet_wxfilename = po.GetArg(3);
TransitionModel trans_model;
AmNnet am_nnet;
{
bool binary_read;
Input ki(nnet_rxfilename, &binary_read);
trans_model.Read(ki.Stream(), binary_read);
am_nnet.Read(ki.Stream(), binary_read);
}
KALDI_ASSERT(minibatch_size > 0);
if (zero_stats) am_nnet.GetNnet().ZeroStats();
double num_examples = 0;
SequentialNnetExampleReader example_reader(examples_rspecifier);
DoBackpropParallel(am_nnet.GetNnet(),
minibatch_size,
&example_reader,
&num_examples,
&(am_nnet.GetNnet()));
{
Output ko(nnet_wxfilename, binary_write);
trans_model.Write(ko.Stream(), binary_write);
am_nnet.Write(ko.Stream(), binary_write);
}
KALDI_LOG << "Finished training, processed " << num_examples
<< " training examples (weighted). Wrote model to "
<< nnet_wxfilename;
return (num_examples == 0 ? 1 : 0);
} catch(const std::exception &e) {
std::cerr << e.what() << '\n';
return -1;
}
}