]>
Commit | Line | Data |
---|---|---|
7974ec9c | 1 | |
2 | **Notes on versioning** | |
3 | ||
4 | ||
5 | ## [Unreleased] | |
518a19d4 VN |
6 | ### Fixes and improvements |
7 | ||
289c4137 FH |
8 | ## [0.9.1](https://github.com/OpenNMT/OpenNMT-py/tree/0.9.1) (2019-06-13) |
9 | * New mechanism for MultiGPU training "1 batch producer / multi batch consumers" | |
10 | resulting in big memory saving when handling huge datasets | |
11 | * New APEX AMP (mixed precision) API | |
12 | * Option to overwrite shards when preprocessing | |
13 | * Small fixes and add-ons | |
14 | ||
87e967dc VN |
15 | ## [0.9.0](https://github.com/OpenNMT/OpenNMT-py/tree/0.9.0) (2019-05-16) |
16 | * Faster vocab building when processing shards (no reloading) | |
17 | * New dataweighting feature | |
18 | * New dropout scheduler. | |
19 | * Small fixes and add-ons | |
20 | ||
518a19d4 VN |
21 | ## [0.8.2](https://github.com/OpenNMT/OpenNMT-py/tree/0.8.2) (2019-02-16) |
22 | * Update documentation and Library example | |
23 | * Revamp args | |
24 | * Bug fixes, save moving average in FP32 | |
25 | * Allow FP32 inference for FP16 models | |
e5aac897 VN |
26 | |
27 | ## [0.8.1](https://github.com/OpenNMT/OpenNMT-py/tree/0.8.1) (2019-02-12) | |
28 | * Update documentation | |
29 | * Random sampling scores fixes | |
30 | * Bug fixes | |
31 | ||
cce528dc VN |
32 | ## [0.8.0](https://github.com/OpenNMT/OpenNMT-py/tree/0.8.0) (2019-02-09) |
33 | * Many fixes and code cleaning thanks @flauted, @guillaumekln | |
34 | * Datasets code refactor (thanks @flauted) you need to r-preeprocess datasets | |
35 | ||
36 | ### New features | |
37 | * FP16 Support: Experimental, using Apex, Checkpoints may break in future version. | |
38 | * Continuous exponential moving average (thanks @francoishernandez, and Marian) | |
39 | * Relative positions encoding (thanks @francoishernanndez, and Google T2T) | |
40 | * Deprecate the old beam search, fast batched beam search supports all options | |
41 | ||
6f11f298 VN |
42 | |
43 | ## [0.7.2](https://github.com/OpenNMT/OpenNMT-py/tree/0.7.2) (2019-01-31) | |
44 | * Many fixes and code cleaning thanks @bpopeters, @flauted, @guillaumekln | |
45 | ||
46 | ### New features | |
47 | * Multilevel fields for better handling of text featuer embeddinggs. | |
48 | ||
49 | ||
5d6f23b2 VN |
50 | ## [0.7.1](https://github.com/OpenNMT/OpenNMT-py/tree/0.7.1) (2019-01-24) |
51 | * Many fixes and code refactoring thanks @bpopeters, @flauted, @guillaumekln | |
7974ec9c | 52 | |
53 | ### New features | |
5d6f23b2 VN |
54 | * Random sampling thanks @daphnei |
55 | * Enable sharding for huge files at translation | |
7974ec9c | 56 | |
f2403462 VN |
57 | ## [0.7.0](https://github.com/OpenNMT/OpenNMT-py/tree/0.7.0) (2019-01-02) |
58 | * Many fixes and code refactoring thanks @benopeters | |
59 | * Migrated to Pytorch 1.0 | |
7974ec9c | 60 | |
37f27333 | 61 | ## [0.6.0](https://github.com/OpenNMT/OpenNMT-py/tree/0.6.0) (2018-11-28) |
62 | * Many fixes and code improvements | |
63 | * New: Ability to load a yml config file. See examples in config folder. | |
64 | ||
9858021f | 65 | ## [0.5.0](https://github.com/OpenNMT/OpenNMT-py/tree/0.5.0) (2018-10-24) |
32af6786 VN |
66 | * Fixed advance n_best beam in translate_batch_fast |
67 | * Fixed remove valid set vocab from total vocab | |
68 | * New: Ability to reset optimizer when using train_from | |
69 | * New: create_vocabulary tool + fix when loading existing vocab. | |
70 | ||
9858021f | 71 | ## [0.4.1](https://github.com/OpenNMT/OpenNMT-py/tree/0.4.1) (2018-10-11) |
70a99a97 VN |
72 | * Fixed preprocessing files names, cleaning intermediary files. |
73 | ||
9858021f | 74 | ## [0.4.0](https://github.com/OpenNMT/OpenNMT-py/tree/0.4.0) (2018-10-08) |
6de42cdb VN |
75 | * Fixed Speech2Text training (thanks Yuntian) |
76 | ||
77 | * Removed -max_shard_size, replaced by -shard_size = number of examples in a shard. | |
78 | Default value = 1M which works fine in most Text dataset cases. (will avoid Ram OOM in most cases) | |
beaf22ba | 79 | |
beaf22ba | 80 | |
9858021f | 81 | ## [0.3.0](https://github.com/OpenNMT/OpenNMT-py/tree/0.3.0) (2018-09-27) |
6de42cdb VN |
82 | * Now requires Pytorch 0.4.1 |
83 | ||
84 | * Multi-node Multi-GPU with Torch Distributed | |
beaf22ba | 85 | |
6de42cdb VN |
86 | New options are: |
87 | -master_ip: ip address of the master node | |
88 | -master_port: port number of th emaster node | |
89 | -world_size = total number of processes to be run (total GPUs accross all nodes) | |
90 | -gpu_ranks = list of indices of processes accross all nodes | |
beaf22ba | 91 | |
6de42cdb | 92 | * gpuid is deprecated |
beaf22ba VN |
93 | See examples in https://github.com/OpenNMT/OpenNMT-py/blob/master/docs/source/FAQ.md |
94 | ||
6de42cdb | 95 | * Fixes to img2text now working |
beaf22ba | 96 | |
6de42cdb | 97 | * New sharding based on number of examples |
beaf22ba | 98 | |
6de42cdb | 99 | * Fixes to avoid 0.4.1 deprecated functions. |
beaf22ba | 100 | |
7974ec9c | 101 | |
9858021f | 102 | ## [0.2.1](https://github.com/OpenNMT/OpenNMT-py/tree/0.2.1) (2018-08-31) |
7974ec9c | 103 | |
104 | ### Fixes and improvements | |
105 | ||
106 | * First compatibility steps with Pytorch 0.4.1 (non breaking) | |
107 | * Fix TranslationServer (when various request try to load the same model at the same time) | |
108 | * Fix StopIteration error (python 3.7) | |
109 | ||
110 | ### New features | |
111 | * Ensemble at inference (thanks @Waino) | |
112 | ||
113 | ## [0.2](https://github.com/OpenNMT/OpenNMT-py/tree/v0.2) (2018-08-28) | |
114 | ||
115 | ### improvements | |
116 | ||
117 | * Compatibility fixes with Pytorch 0.4 / Torchtext 0.3 | |
118 | * Multi-GPU based on Torch Distributed | |
119 | * Average Attention Network (AAN) for the Transformer (thanks @francoishernandez ) | |
120 | * New fast beam search (see -fast in translate.py) (thanks @guillaumekln) | |
121 | * Sparse attention / sparsemax (thanks to @bpopeters) | |
122 | * Refactoring of many parts of the code base: | |
123 | - change from -epoch to -train_steps -valid_steps (see opts.py) | |
124 | - reorg of the logic train => train_multi / train_single => trainer | |
125 | * Many fixes / improvements in the translationserver (thanks @pltrdy @francoishernandez) | |
126 | * fix BPTT | |
127 | ||
128 | ## [0.1](https://github.com/OpenNMT/OpenNMT-py/tree/v0.1) (2018-06-08) | |
129 | ||
130 | ### First and Last Release using Pytorch 0.3.x | |
131 | ||
132 |