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
|
# AmuNMT
[](https://gitter.im/amunmt/amunmt?utm_source=badge&utm_medium=badge&utm_campaign=pr-badge&utm_content=badge)
[](http://vali.inf.ed.ac.uk/jenkins/job/amunmt_compilation_cuda/)
[](http://vali.inf.ed.ac.uk/jenkins/job/amunmt_compilation_cpu/)
A C++ inference engine for Neural Machine Translation (NMT) models trained with Theano-based scripts from
Nematus (https://github.com/rsennrich/nematus) or DL4MT (https://github.com/nyu-dl/dl4mt-tutorial)
If you use this, please cite:
Marcin Junczys-Dowmunt, Tomasz Dwojak, Hieu Hoang (2016). Is Neural Machine Translation Ready for Deployment? A Case Study on 30 Translation Directions (https://arxiv.org/abs/1610.01108)
## Recommended for GPU version:
Tested on Ubuntu 14.04 LTS
* CMake 3.5.1 (due to CUDA related bugs in earlier versions)
* GCC/G++ 4.9
* Boost 1.54
* CUDA 7.5
Tested on Ubuntu 16.04 LTS
* CMake 3.5.1 (due to CUDA related bugs in earlier versions)
* GCC/G++ 5.4
* Boost 1.61
* CUDA 8.0
Also compiles the CPU version.
## Recommended for CPU version:
The CPU-only version will automatically be compiled if CUDA cannot be detected by CMAKE. Tested on different machines and distributions:
* CMake 3.5.1
* The CPU version should be a lot more forgiving concerning GCC/G++ or Boost versions.
## Git repository set-up (branch marian-integration only)
Branch _marian-integration_ now contains marian as a sub-module.
After the initial clone or checkout of this branch, run the following command
```bash
git submodule update --init --recursive
```
If you want git to always update the submodules at git checkout, git pull, git merge, etc. (it won't by default, wich often causes confusion), copy the script ```git-hooks/post-rewrite``` into ```./git/hooks``` and make it executable.
## Compilation
The project is a standard Cmake out-of-source build:
mkdir build
cd build
cmake ..
make -j
If you want to compile only CPU version on a machine with CUDA, add `-DCUDA=OFF` flag:
cmake -DCUDA=OFF ..
## Vocabulary files
Vocabulary files (and all other config files) in AmuNMT are by default YAML files. AmuNMT also reads gzipped yml.gz files.
* Vocabulary files from models trained with Nematus can be used directly as JSON is a proper subset of YAML.
* Vocabularies for models trained with DL4MT (\*.pkl extension) need to be converted to JSON/YAML with either of the two scripts below:
```
python scripts/pkl2json.py vocab.en.pkl > vocab.json
python scripts/pkl2yaml.py vocab.en.pkl > vocab.yml
```
## Running AmuNMT
./bin/amun -c config.yml <<< "This is a test ."
## Configuration files
An example configuration:
# Paths are relative to config file location
relative-paths: yes
# performance settings
beam-size: 12
devices: [0]
normalize: yes
gpu-threads: 1
# scorer configuration
scorers:
F0:
path: model.en-de.npz
type: Nematus
# scorer weights
weights:
F0: 1.0
# vocabularies
source-vocab: vocab.en.yml.gz
target-vocab: vocab.de.yml.gz
## BPE Support
AmuNMT has integrated support for [BPE encoding](https://github.com/rsennrich/subword-nmt). There are two option `bpe` and `debpe`. The `bpe` option receives a path to a file with BPE codes (here `bpe.codes`). To turn on desegmentation on the ouput, set `debpe` to `true`, e.g.
bpe: bpe.codes
debpe: true
## Python Bindings
Python bindings allow to run AmuNMT decoder in python scripts. The compilation of the bindings requires `python-dev` package. To compile the bindings run:
```
make python
```
The Python bindings consist of 3 functions: `init`, `translate`, and `shutdown`:
```python
import libamunmt
libamunmt.init('-c config.yml')
print libamunmt.translate(['this is a little test .'])
libamunmt.shutdown()
```
The `init` function init the decoder and the syntax is the same as in command line. The `translate`
function takes a list of sentences to translate. For real-world example, see the `scripts/amunmt_server.py`
script, which uses python bindings to run REST server.
The function `shutdown` is needed (and should be called at the end of your script) to avoid runtime errors
due to the random order in which objects may be deallocated if they are not explicitly destroyed.
## Using GPU/CPU threads
AmuNMT can use GPUs, CPUs, or both, to distribute translation of different sentences.
**However, it is unlikely that CPUs used together with GPUs yield any performance improvement.
It is probably better to only use the GPU if one or more are available.**
cpu-threads: 8
gpu-threads: 2
devices: [0, 1]
The setting above uses 8 CPU threads and 4 GPU threads (2 GPUs x 2 threads). The `gpu-threads` and `devices` options are only available when AmuNMT has been compiled with CUDA support. Multiple GPU threads can be used to increase GPU saturation, but will likely not result in a large performance boost. By default, `gpu-threads` is set to `1` and `cpu-threads` to `0` if CUDA is available. Otherwise `cpu-threads` is set to `1`. To disable the GPU set `gpu-threads` to `0`. Setting both `gpu-threads` and `cpu-threads` to `0` will result in an exception.
## Example usage
* [Data and systems for our winning system in the WMT 2016 Shared Task on Automatic Post-Editing](https://github.com/emjotde/amunmt/wiki/AmuNMT-for-Automatic-Post-Editing)
## Acknowledgements
The development of Amunmt received funding from the European Union's Horizon 2020 Research and Innovation Programme under grant agreements 688139 (SUMMA; 2016-2019) and 645487 (Modern MT; 2015-2017) and the Amazon Academic Research Awards program.
|
