mlm-pipeline

mlm-pipeline is a cloud architecture that preprocesses the masked language model (mlm).

In NLP, a masked Languge Model (MLM) such as BERT, XLM, RoBERTa, and ALBERT, pretraining the sentence's input with [MASK] is a state-of-a-art.

Input Text: the man jumped up , put his basket on phil ##am ##mon ' s head Original Masked Input: [MASK] man [MASK] up , put his [MASK] on phil [MASK] ##mon ' s head

However, the preprocessing process of tokenizing and masking a few hundred GB of large text takes a lot of time with a single node. We use a multi-node architecture that distributes preprocessing through the cloud architecture's pipeline design with pull-push pattern.

• ventilator : Read large text and deliver message to zmq's queue. ventilator is a single node.
• worker : 1) BERT Tokenizer, 2) Create Masked on sentences, 3) push preprocessed tfrecord to S3
• worker controller : using Terraform, Ansible, we can control all ec2s and dynamic provisioning ec2 on AWS.

Usage

prepare wiki dump data

(If you don't use this wiki data, you can cancel this step). extract the text with WikiExtractor.py. It took about an hour using 96 core ec2.

wget https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2
bzip2 -dk enwiki-latest-pages-articles.xml.bz2

python3 WikiExtractor.py -o \
  ../output --processes 80 \
  ../enwiki-latest-pages-articles.xml

dynamic provisioning EC2(worker) on AWS

• Installing Terraform
• Installation Guide Ansible

(run on your local pc)
chmod +x init.sh
./init.sh
git clone https://github.com/graykode/mlm-pipeline

export AWS_ACCESS_KEY_ID='xxxxxxx'
export AWS_SECRET_ACCESS_KEY='xxxxxx'

cd worker_controller/terraform
terraform init

change some variable in variables.tf

• region, zone, number_of_worker , client_instance_type, volume_size, client_subnet, client_security_groups, default_keypair_name
• you must open ventilation port when create client_security_groups.

Then run below:

(run on your local pc)
terraform apply

or if you want to destroy all, type terraform destroy

command to all ec2 nodes in one time.

(run on your local pc)
cd ../ansible

# ping test
ansible-playbook -i ./inventory/ec2.py \
      --limit "tag_type_worker" \
      -u ubuntu \
      --private-key ~/.ssh/SoRT.pem ping.yaml
      
# install python packagement(ex tensorflow, boto, zmq, ..)
ansible-playbook -i ./inventory/ec2.py \
      --limit "tag_type_worker" \
      -u ubuntu \
      --private-key ~/.ssh/SoRT.pem init.yaml \
      --extra-vars "aws_access_key_id=<key_id> aws_secret_access_key=<access_key>" -vvvv

aws_access_key_id and aws_secret_access_key will be in environment variable (/etc/environment) to using boto s3. change as your <key_id>, <access_key>.

ventilation setting

(in ventilation ec2)
wget https://raw.githubusercontent.com/graykode/mlm-pipeline/master/init.sh
# init shell for ventilator
sudo apt update && sudo apt install -y python3 && \
      sudo apt install -y python3-pip && \
      pip3 install zmq

Start in workers and ventilation order

(run on your local pc)
ansible-playbook -i ./inventory/ec2.py \
      --limit "tag_type_worker" \
      -u ubuntu \
      --private-key ~/.ssh/SoRT.pem working.yaml \
      --extra-vars "bucket_name=<bucket_name> vserver=<ventilator_ip>"
      
(in ventilation ec2)
python3 ventilator.py \
  --data 'data folder path' \
  --vport 5557 \
  --time 0.88  

License

MIT

Author

• Tae Hwan Jung(Jeff Jung) @graykode, Kyung Hee Univ CE(Undergraduate).
• Author Email : [email protected]