The 1st Workshop on Automatic Simultaneous Translation

Challenges, Recent Advances, and Future Directions

Workshop at ACL 2020, Seattle, July 10, 2020
Contact: autosimtrans.workshop@gmail.com or twitter.com/autosimtrans

A. pack your “trained model + code” into docker

- English

1. for authers: build docker with your own model and code
2. for users: test a model in docker

- 中文

B. train in docker from stratch

- English

1. train a transformer-tf in docker

1. For authers who want to build docker images with your own models

  1. Build docker image, from a Dockerfile, conda_specfile.txt and pip_requirement.txt will be used in Dockerfile 
    conda activate <myenv>				# activate your virtual env
conda list --explicit > conda_specfile.txt 	# export packages installed via conda
pip freeze > pip_requirement.txt		# export packages installed via pip
  2. cd to your working dir, download this Dockerfile(named as Dockerfile_pt13.gpu) 
    wget https://raw.githubusercontent.com/autosimtrans/autosimtrans.github.io/master/sp/Dockerfile_pt13.gpu
  3. Copy your selected code+model into a folder in this working dir, noted as <code folder>. Open the Dockerfile Dockerfile_pt13.gpu, and replace OpenNMT-py with the <code folder> of your own code and model (the reason of copying code+model is that CP in dockerfile only supports relative path). Update your own CUDA/CuDnn version in Dockerfile if necessary.
  4. Build docker image 
    docker build -t <ImageName:Tag> -f Dockerfile_pt13.gpu .
  5. Create local folder, which will mount to docker container, and need write access 
    mkdir decode && chmod 777 decode
  6. After image finished, test your docker image by running inference 
    nvidia-docker run -it -v <abs_path_to_your_data_dir>:/data -v <abs_path_to_your_local_decode_result_dir>:/decode <ImageName:Tag> bash
  7. Now you are in the docker container with bash end_point, just run the translate script, for example: 
    python translate.py -model multi30k_model_step_100000.pt -src /data/wmt16-multi30k/test2016.en.atok  -tgt /data/wmt16-multi30k/test2016.de.atok  -replace_unk -verbose -output /decode/multi30k.test.pred.atok
  8. Then you can check the translation result in either <abs_path_to_your_local_decode_result_dir>(local) or /decode (in docker container), test your BLEU: 
    # in docker conatiner
perl tools/multi-bleu.perl /data/wmt16-multi30k/test2016.de.atok < /decode/multi30k.test.pred.atok
# result:
# BLEU = 35.50, 65.9/41.8/28.8/20.0 (BP=1.000, ratio=1.007, hyp_len=12323, ref_len=12242)
  9. Push your image to docker hub or save your image into a tar file:
    • sign up a DockerHub account, create a new repo, push your image to docker hub 
       docker tag <local-image:loacl-tagname> <dockerhub_username/new-repo:repo-tagname>
 docker login
 docker push <dockerhub_username/new-repo:repo-tagname>
    • save your image to a tar file, then you can send it to anyone: 
       # save image to `tar` file
 docker save -o <imageFile>.tar <local-image:loacl-tagname>
 # load `tar` file to image
 docker load < <imageFile>.tar

2. For users who want to use docker images(covering code+model)

  1. cd to your working dir, and download data (for the case of blind test) 
    wget https://github.com/autosimtrans/autosimtrans.github.io/raw/master/sp/wmtdata.tar.gz
tar -zxvf wmtdata.tar.gz
  2. Create local folder, which will mount to docker container, and need write access 
    mkdir decode && chmod 777 decode
  3. Run docker(nvidia-docker is used as GPU version) 
    sudo nvidia-docker run -it -v <abs_path_to_your_data_dir>:/data -v <abs_path_to_your_local_decode_result_dir>:/decode kaiboliu/onmt-py_en2de:torch1.3-gpu bash
  4. Now you are in the docker container with bash end_point, just run the translate script: 
    python translate.py -model multi30k_model_step_100000.pt -src /data/wmt16-multi30k/test2016.en.atok -replace_unk -verbose -output /decode/multi30k.test.pred.atok
  5. Then you can check the translation result in either <abs path to your local decode result dir>(local) or /decode (in docker container), test your BLEU: 
    # in docker conatiner
perl tools/multi-bleu.perl /data/wmt16-multi30k/test2016.de.atok < /decode/multi30k.test.pred.atok
# result:
# BLEU = 35.50, 65.9/41.8/28.8/20.0 (BP=1.000, ratio=1.007, hyp_len=12323, ref_len=12242)

Train a transformer-tf in docker

1. Run your model in Docker

  1. You can walk through Step 1-12 from strach to understand the whole pipeline
  2. If you have your own code/model ready, you can skip Step 3 and Step 6-9, just copy your data to the container by 
    docker cp <local_data_path>:[containerID]:<docker_data_path>

2. Training a TensorFlow-based Transformer Model from Scratch in Docker

We will dive head-first into training a transformer model from scratch using a TensorFlow GPU Docker image.

Step 1) Launch TensorFlow GPU Docker Container

Using Docker allows us to spin up a fully contained environment for our training needs. We always recommend using Docker, as it allows ultimate flexibility (and forgiveness) in our training environment. To begin we will open a terminal window and enter the following command to launch our NVIDIA CUDA powered container.

nvidia-docker run -it -p 6007:6006 -v /data:/datasets tensorflow/tensorflow:1.15.0-gpu-py3 bash

Note: A quick description about the key parameters of the above command (if you’re unfamiliar with Docker).

Docker Syntax Description
nvidia-docker run The ‘docker run’ command specifies the container from. Note in this case we use ‘nvidia-docker run’ to utilize CUDA powered NVIDIA GPUs.
-p 6007:6006 Expose port 6007 [HOST:CONTAINER] for Tensorboard, type localhost:6007 in your browser to view Tensorboard
-v data:/datasets Volume tag, This shares the folder ‘/data’ on the host machine to the /datasets folder in the container. Also explained as: -v /[host folder]:/[container folder]
tensorflow/tensorflow:nightly-gpu This the docker image that will run. The Docker Hub format is [PUBLISHER]/[IMAGE REPO]:[IMAGE TAG]
hint

You can also use the framework combination named Deepo

Step 2) Install git

This may be necessary if you are running a fresh docker container.

apt-get install git

Step 3) Download ‘Models’ (Transformer Codebase is included here)

In case you do not have the latest up-to-date codebase for the models, the transformer network model is included here and the devs tend to update it quite frequently.

mkdir /tf-1.5; cd /tf-1.5
git clone https://github.com/tensorflow/models.git

Step 4) Install Requirements for TensorFlow Models

As a necessary step,this will install the python package requirements for training TensorFlow models.

# cd to the models dir
pip install --user -r official/requirements.txt

Step 5) Export Pythonpath

Export PYTHONPATH to the folder where the models folder are located on your machine. The command below references where the models are located on our system. Be sure to replace the /transformer/models syntax with the data path to the folder where you stored/downloaded your models.

export PYTHONPATH="$PYTHONPATH:/tf-1.5/models"

Step 6) Download and Preprocess the Dataset

The data_download.py command will download and preprocess the training and evaluation WMT datasets. Upon download and extraction, the training data is used to generate for what we will use as VOCAB_FILE variables. Effectively, the eval and training strings are tokenized, and the results are processed and saved as TFRecords.

NOTE: (per the official requirements): 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.

cd /tf-1.5/models/official/transformer
python data_download.py --data_dir=/datasets/transformer

Note: you can skip step 7-9 if you already finished training your model, just test it in the docker from Step 10

Step 7) Set Training Variables for the Transformer Model

‘PARAM_SET’

This specifies what model to train. big or base

PARAM_SET=base
‘DATA_DIR’

This variable should be set to where the training data is located.

DATA_DIR=/datasets/transformer
‘MODEL_DIR’

This variable specifies the model location based on what model is specified in the PARAM_SET variable

MODEL_DIR=/transformer/model_$PARAM_SET
‘VOCAB_FILE’

This variable expresses where the location of the preprocessed vocab files are located.

VOCAB_FILE=$DATA_DIR/vocab.ende.32768
‘EXPORT_DIR’ Export trained transformer model

This will specify the location when/where you export the model in Tensorflow SavedModel format. This is done when using the flag export_dir when training in step 8.

EXPORT_DIR=/transformer/saved_model

PARAM_SET=base
DATA_DIR=/datasets/transformer
MODEL_DIR=/transformer/model_$PARAM_SET
VOCAB_FILE=$DATA_DIR/vocab.ende.32768
EXPORT_DIR=/transformer/saved_model

Step 8) Train the Transformer Model

The following command ‘python transformer_main.py’ will train the transformer for a total of 260,000 steps. See how the flags are set up to reference the variables you set in the previous steps. You can train for less than 260,000 steps, it’s up to you.

NOTE: This will take a long time to train depending on your GPU resources. The official TensorFlow transformer model is under constant development, be sure to check periodically on their github for any latest optimizations and techniques to reduce training times.

python transformer_main.py --data_dir=$DATA_DIR --model_dir=$MODEL_DIR --vocab_file=$VOCAB_FILE --param_set=$PARAM_SET --bleu_source=$DATA_DIR/newstest2014.en --bleu_ref=$DATA_DIR/newstest2014.de --train_steps=260000 --steps_between_evals=1000 --export_dir=$EXPORT_DIR

Step 9) View Transformer Model Results in Tensorboard

As we noted earlier, we can check the status of training in the Tensorboard GUI. To check in real time, run the following command in a separate terminal (or TensorFlow container), and type localhost:6007 in your browser to view Tensorboard. You can also wait until training is complete to use the current container.

tensorboard --logdir=$MODEL_DIR

Step 10) Test the Trained Transformer Network (Translate English to German)

Now we’ve trained our transformer network, let’s enjoy the fruits of our labor using translate.py! In the command below, replace the text “hello world” with desired text to translate

python translate.py --model_dir=$MODEL_DIR --vocab_file=$VOCAB_FILE \
--param_set=$PARAM_SET --text="hello world"

Step 11) save your container to a image

  1. Create an account and a repo on Docker Hub, named [UserID] and [Repo]
  2. Get the [containerID] you want by docker ps -l
  3. construct the image locally 
    docker commit [containerID] [UserID]/[Repo]
  4. check the image 
    docker images
  5. Upload your image to Docker Hub, you may be prompted with username and password of Docker Hub 
    docker push [UserID]/[Repo]