Transformer was proposed in 2017 and designed to process sequential data. It is adopted mainly in the field of natural language processing(NLP), for tasks like machine translation or text summarization. Unlike traditional recurrent neural network(RNN) which processes data in order, Transformer adopts attention mechanism and improve the parallelism, therefore reduced training times and made training on larger datasets possible. Since Transformer model was introduced, it has been used to tackle many problems in NLP and derives many network models, such as BERT(Bidirectional Encoder Representations from Transformers) and GPT(Generative Pre-trained Transformer).
[Paper](https://arxiv.org/abs/1706.03762): Ashish Vaswani, Noam Shazeer, Niki Parmar, JakobUszkoreit, Llion Jones, Aidan N Gomez, Ł ukaszKaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS 2017, pages 5998–6008.
# [Model Architecture](#contents)
Specifically, Transformer contains six encoder modules and six decoder modules. Each encoder module consists of a self-attention layer and a feed forward layer, each decoder module consists of a self-attention layer, a encoder-decoder-attention layer and a feed forward layer.
# [Dataset](#contents)
- *WMT Englis-German* for training.
- *WMT newstest2014* for evaluation.
# [Environment Requirements](#contents)
- Hardware（Ascend）
- Prepare hardware environment with Ascend processor. If you want to try Ascend , please send the [application form](https://obs-9be7.obs.cn-east-2.myhuaweicloud.com/file/other/Ascend%20Model%20Zoo%E4%BD%93%E9%AA%8C%E8%B5%84%E6%BA%90%E7%94%B3%E8%AF%B7%E8%A1%A8.docx) to ascend@huawei.com. Once approved, you can get the resources.
- You may use this [shell script](https://github.com/tensorflow/nmt/blob/master/nmt/scripts/wmt16_en_de.sh) to download and preprocess WMT English-German dataset. Assuming you get the following files:
- train.tok.clean.bpe.32000.en
- train.tok.clean.bpe.32000.de
- vocab.bpe.32000
- newstest2014.tok.bpe.32000.en
- newstest2014.tok.bpe.32000.de
- newstest2014.tok.de
- Convert the original data to mindrecord for training:
- Set options in `config.py`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/tutorial/zh-CN/master/use/data_preparation/loading_the_datasets.html#mindspore) for more information about dataset.
- Run `run_standalone_train_ascend.sh` for non-distributed training of Transformer model.
``` bash
sh scripts/run_standalone_train_ascend.sh DEVICE_ID EPOCH_SIZE DATA_PATH
```
- Run `run_distribute_train_ascend.sh` for distributed training of Transformer model.
``` bash
sh scripts/run_distribute_train_ascend.sh DEVICE_NUM EPOCH_SIZE DATA_PATH RANK_TABLE_FILE
```
## [Evaluation Process](#contents)
- Set options in `eval_config.py`. Make sure the 'data_file', 'model_file' and 'output_file' are set to your own path.
- Run `eval.py` for evaluation of Transformer model.
```bash
python eval.py
```
- Run `process_output.sh` to process the output token ids to get the real translation results.
```bash
sh scripts/process_output.sh REF_DATA EVAL_OUTPUT VOCAB_FILE
```
You will get two files, REF_DATA.forbleu and EVAL_OUTPUT.forbleu, for BLEU score calculation.
- Calculate BLEU score, you may use this [perl script](https://github.com/moses-smt/mosesdecoder/blob/master/scripts/generic/multi-bleu.perl) and run following command to get the BLEU score.
Some seeds have already been set in train.py to avoid the randomness of dataset shuffle and weight initialization. If you want to disable dropout, please set the corresponding dropout_prob parameter to 0 in src/config.py.
# [ModelZoo Homepage](#contents)
Please check the official [homepage](https://gitee.com/mindspore/mindspore/tree/master/model_zoo).
