Chainer-based Python implementation of Transformer, an attention-based seq2seq model without convolution and recurrence.
If you want to see the architecture, please see net.py.
See "Attention Is All You Need", Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin, arxiv, 2017.
This repository is partly derived from my convolutional seq2seq repo, which is also derived from Chainer's official seq2seq example.
- Python 3.6.0+
- Chainer 2.0.0+
- numpy 1.12.1+
- cupy 1.0.0+ (if using gpu)
- nltk
- progressbar
- (You can install all through
pip
) - and their dependencies
You can use any parallel corpus.
For example, run
sh download_wmt.sh
which downloads and decompresses training dataset and development dataset from WMT/europal into your current directory. These files and their paths are set in training script train.py
as default.
PYTHONIOENCODING=utf-8 python -u train.py -g=0 -i DATA_DIR -o SAVE_DIR
During training, logs for loss, perplexity, word accuracy and time are printed at a certain internval, in addition to validation tests (perplexity and BLEU for generation) every half epoch. And also, generation test is performed and printed for checking training progress.
Some of them is as follows:
-g
: your gpu id. If cpu, set-1
.-i DATA_DIR
,-s SOURCE
,-t TARGET
,-svalid SVALID
,-tvalid TVALID
:
DATA_DIR
directory needs to include a pair of training datasetSOURCE
andTARGET
with a pair of validation datasetSVALID
andTVALID
. Each pair should be parallell corpus with line-by-line sentence alignment.-o SAVE_DIR
: JSON log report file and a model snapshot will be saved inSAVE_DIR
directory (if it does not exist, it will be automatically made).-e
: max epochs of training corpus.-b
: minibatch size.-u
: size of units and word embeddings.-l
: number of layers in both the encoder and the decoder.--source-vocab
: max size of vocabulary set of source language--target-vocab
: max size of vocabulary set of target language
Please see the others by python train.py -h
.
This repository does not aim for complete validation of results in the paper, so I have not eagerly confirmed validity of performance. But, I expect my implementation is almost compatible with a model described in the paper. Some differences where I am aware are as follows:
- Optimization/training strategy. Detailed information about batchsize, parameter initialization, etc. is unclear in the paper. Additionally, the learning rate proposed in the paper may work only with a large batchsize (e.g. 4000) for deep layer nets. I changed warmup_step to 32000 from 4000, though there is room for improvement. I also changed
relu
intoleaky relu
in feedforward net layers for easy gradient propagation. - Vocabulary set, dataset, preprocessing and evaluation. This repo uses a common word-based tokenization, although the paper uses byte-pair encoding. Size of token set also differs. Evaluation (validation) is little unfair and incompatible with one in the paper, e.g., even validation set replaces unknown words to a single "unk" token.
- Beam search is unused in BLEU calculation.
- Model size. The setting of a model in this repo is one of "base model" in the paper, although you can modify some lines for using "big model".
- This code follows some settings used in tensor2tensor repository, which includes a Transformer model. For example, positional encoding used in the repository seems to differ from one in the paper. This code follows the former one.