Training is accomplished using the DistributionStrategies API. (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/distribute/README.md)
The appropriate distribution strategy is chosen based on the `--num_gpus` flag. By default this flag is one if TensorFlow is compiled with CUDA, and zero otherwise.
num_gpus:
+ 0: Use OneDeviceStrategy and train on CPU.
+ 1: Use OneDeviceStrategy and train on GPU.
+ 2+: Use MirroredStrategy (data parallelism) to distribute a batch between devices.
This is an implementation of the Transformer translation model as described in the [Attention is All You Need](https://arxiv.org/abs/1706.03762) paper. Based on the code provided by the authors: [Transformer code](https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/models/transformer.py) from [Tensor2Tensor](https://github.com/tensorflow/tensor2tensor).
Transformer is a neural network architecture that solves sequence to sequence problems using attention mechanisms. Unlike traditional neural seq2seq models, Transformer does not involve recurrent connections. The attention mechanism learns dependencies between tokens in two sequences. Since attention weights apply to all tokens in the sequences, the Tranformer model is able to easily capture long-distance depedencies.
Transformer's overall structure follows the standard encoder-decoder pattern. The encoder uses self-attention to compute a representation of the input sequence. The decoder generates the output sequence one token at a time, taking the encoder output and previous decoder-outputted tokens as inputs.
The model also applies embeddings on the input and output tokens, and adds a constant positional encoding. The positional encoding adds information about the position of each token.
Below are the commands for running the Transformer model. See the [Detailed instrutions](#detailed-instructions) for more details on running the model.
```
PARAMS=big
DATA_DIR=$HOME/transformer/data
MODEL_DIR=$HOME/transformer/model_$PARAMS
# Download training/evaluation datasets
python data_download.py --data_dir=$DATA_DIR
# Train the model for 10 epochs, and evaluate after every epoch.
Currently, both big and base params run on a single GPU. The measurements below
are reported from running the model on a P100 GPU.
Params | batches/sec | batches per epoch | time per epoch
--- | --- | --- | ---
base | 4.8 | 83244 | 4 hr
big | 1.1 | 41365 | 10 hr
### Evaluation results
Below are the case-insensitive BLEU scores after 10 epochs.
Params | Score
--- | --- |
base | 27.7
big | 28.9
## Detailed instructions
0. ### Export variables (optional)
Export the following variables, or modify the values in each of the snippets below:
```
PARAMS=big
DATA_DIR=$HOME/transformer/data
MODEL_DIR=$HOME/transformer/model_$PARAMS
```
1. ### Download and preprocess datasets
[data_download.py](data_download.py) downloads and preprocesses the training and evaluation WMT datasets. After the data is downloaded and extracted, the training data is used to generate a vocabulary of subtokens. The evaluation and training strings are tokenized, and the resulting data is sharded, shuffled, and saved as TFRecords.
1.75GB of compressed data will be downloaded. In total, the raw files (compressed, extracted, and combined files) take up 8.4GB of disk space. The resulting TFRecord and vocabulary files are 722MB. The script takes around 40 minutes to run, with the bulk of the time spent downloading and ~15 minutes spent on preprocessing.
Command to run:
```
python data_download.py --data_dir=$DATA_DIR
```
Arguments:
*`--data_dir`: Path where the preprocessed TFRecord data, and vocab file will be saved.
* Use the `--help` or `-h` flag to get a full list of possible arguments.
2. ### Model training and evaluation
[transformer_main.py](transformer_main.py) creates a Transformer model, and trains it using Tensorflow Estimator.
*`--data_dir`: This should be set to the same directory given to the `data_download`'s `data_dir` argument.
*`--model_dir`: Directory to save Transformer model training checkpoints.
*`--params`: Parameter set to use when creating and training the model. Options are `base` and `big` (default).
* Use the `--help` or `-h` flag to get a full list of possible arguments.
#### Customizing training schedule
By default, the model will train for 10 epochs, and evaluate after every epoch. The training schedule may be defined through the flags:
* Training with epochs (default):
*`--train_epochs`: The total number of complete passes to make through the dataset
*`--epochs_between_eval`: The number of epochs to train between evaluations.
* Training with steps:
*`--train_steps`: sets the total number of training steps to run.
*`--steps_between_eval`: Number of training steps to run between evaluations.
Only one of `train_epochs` or `train_steps` may be set. Since the default option is to evaluate the model after training for an epoch, it may take 4 or more hours between model evaluations. To get more frequent evaluations, use the flags `--train_steps=250000 --steps_between_eval=1000`.
Note: At the beginning of each training session, the training dataset is reloaded and shuffled. Stopping the training before completing an epoch may result in worse model quality, due to the chance that some examples may be seen more than others. Therefore, it is recommended to use epochs when the model quality is important.
#### Compute BLEU score during model evaluation
Use these flags to compute the BLEU when the model evaluates:
*`--bleu_source`: Path to file containing text to translate.
*`--bleu_ref`: Path to file containing the reference translation.
*`--bleu_threshold`: Train until the BLEU score reaches this lower bound. This setting overrides the `--train_steps` and `--train_epochs` flags.
The test source and reference files located in the `test_data` directory are extracted from the preprocessed dataset from the [NMT Seq2Seq tutorial](https://google.github.io/seq2seq/nmt/#download-data).
When running `transformer_main.py`, use the flags: `--bleu_source=test_data/newstest2014.en --bleu_ref=test_data/newstest2014.de`
#### Tensorboard
Training and evaluation metrics (loss, accuracy, approximate BLEU score, etc.) are logged, and can be displayed in the browser using Tensorboard.
```
tensorboard --logdir=$MODEL_DIR
```
The values are displayed at [localhost:6006](localhost:6006).
3. ### Translate using the model
[translate.py](translate.py) contains the script to use the trained model to translate input text or file. Each line in the file is translated separately.
*`--translation`: Path to file containing generated translations.
*`--reference`: Path to file containing reference translations.
* Use the `--help` or `-h` flag to get a full list of possible arguments.
## Implementation overview
A brief look at each component in the code:
### Model Definition
The [model](model) subdirectory contains the implementation of the Transformer model. The following files define the Transformer model and its layers:
*[transformer.py](model/transformer.py): Defines the transformer model and its encoder/decoder layer stacks.
*[embedding_layer.py](model/embedding_layer.py): Contains the layer that calculates the embeddings. The embedding weights are also used to calculate the pre-softmax probabilities from the decoder output.
*[attention_layer.py](model/attention_layer.py): Defines the multi-headed and self attention layers that are used in the encoder/decoder stacks.
*[ffn_layer.py](model/ffn_layer.py): Defines the feedforward network that is used in the encoder/decoder stacks. The network is composed of 2 fully connected layers.
Other files:
*[beam_search.py](model/beam_search.py) contains the beam search implementation, which is used during model inference to find high scoring translations.
*[model_params.py](model/model_params.py) contains the parameters used for the big and base models.
*[model_utils.py](model/model_utils.py) defines some helper functions used in the model (calculating padding, bias, etc.).
### Model Estimator
[transformer_main.py](model/transformer.py) creates an `Estimator` to train and evaluate the model.
Helper functions:
*[utils/dataset.py](utils/dataset.py): contains functions for creating a `dataset` that is passed to the `Estimator`.
*[utils/metrics.py](utils/metrics.py): defines metrics functions used by the `Estimator` to evaluate the
### Other scripts
Aside from the main file to train the Transformer model, we provide other scripts for using the model or downloading the data:
#### Data download and preprocessing
[data_download.py](data_download.py) downloads and extracts data, then uses `Subtokenizer` to tokenize strings into arrays of int IDs. The int arrays are converted to `tf.Examples` and saved in the `tf.RecordDataset` format.
The data is downloaded from the Workshop of Machine Transtion (WMT) [news translation task](http://www.statmt.org/wmt17/translation-task.html). The following datasets are used:
* Europarl v7
* Common Crawl corpus
* News Commentary v12
See the [download section](http://www.statmt.org/wmt17/translation-task.html#download) to explore the raw datasets. The parameters in this model are tuned to fit the English-German translation data, so the EN-DE texts are extracted from the downloaded compressed files.
The text is transformed into arrays of integer IDs using the `Subtokenizer` defined in [`utils/tokenizer.py`](util/tokenizer.py). During initialization of the `Subtokenizer`, the raw training data is used to generate a vocabulary list containing common subtokens.
The target vocabulary size of the WMT dataset is 32,768. The set of subtokens is found through binary search on the minimum number of times a subtoken appears in the data. The actual vocabulary size is 33,708, and is stored in a 324kB file.
#### Translation
Translation is defined in [translate.py](translate.py). First, `Subtokenizer` tokenizes the input. The vocabulary file is the same used to tokenize the training/eval files. Next, beam search is used to find the combination of tokens that maximizes the probability outputted by the model decoder. The tokens are then converted back to strings with `Subtokenizer`.
#### BLEU computation
[compute_bleu.py](compute_bleu.py): Implementation from [https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/utils/bleu_hook.py](https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/utils/bleu_hook.py).
### Test dataset
The [newstest2014 files](test_data) are extracted from the [NMT Seq2Seq tutorial](https://google.github.io/seq2seq/nmt/#download-data). The raw text files are converted from the SGM format of the [WMT 2016](http://www.statmt.org/wmt16/translation-task.html) test sets.
## Term definitions
**Steps / Epochs**:
* Step: unit for processing a single batch of data
* Epoch: a complete run through the dataset
Example: Consider a training a dataset with 100 examples that is divided into 20 batches with 5 examples per batch. A single training step trains the model on one batch. After 20 training steps, the model will have trained on every batch in the dataset, or one epoch.
**Subtoken**: Words are referred to as tokens, and parts of words are referred to as 'subtokens'. For example, the word 'inclined' may be split into `['incline', 'd_']`. The '\_' indicates the end of the token. The subtoken vocabulary list is guaranteed to contain the alphabet (including numbers and special characters), so all words can be tokenized.
