Awesome

flair-on-gcp

Please note: this repository is not affiliated with Google.

This repository adds examples on how to train Flair text classification model on Google Cloud Platform (GCP) AI Platform.

It also covers how to deploy a pre-trained Flair-model on GCP Cloud Run using Cloud Build and Cloud Registry to serve predictions.

Please note: this repository is only meant as an example, and is WIP.

Getting started
Training guide
Deployment guide

Getting started

Google Cloud SDK

It is recommended to install the Google Cloud SDK so that you can interact with Google Cloud services from your local command line interface (CLI).

You can find instructions to install the Google Cloud SDK here.

Access management

Please review whether you have proper access rights to the appropriate services by checking the IAM documentation for each service.

To get started:

Training

Deployment

Training

It is possible to submit training jobs on GCP AI platform by making use of Docker Containers.

This allows us:

To make use of open source frameworks relatively easily.
To scale up with computational requirements.
To take a "hands off" approach in comparison to spinning up AI platform notebooks.

The training folder contains a Docker file with requirements, and a Python script to train a text classification model facilitated by Flair.

Overview

The Docker image installs the required libraries and creates the following folder structure:

/root/trainer
         ├── text-classification-training.py
         ├── data
         |     ├── test.csv (optional)
         |     ├── dev.csv (optional)
         │     └── train.csv
         └── checkpoint
               ├── final-model.pt
               ├── best-model.pt
               └── training.log

The text classification script also discusses the parameters more in-depth.

The gist is, that at training time, the Docker container executes the text classification script which:

Parses the supplied arguments.
Handles ingress of test.csv / train.csv / dev.csv files from the GCS bucket to the container.
Initiates a Flair text-classification loop with the parsed arguments.
Handles egress of the trained models and logging files to the GCS bucket.

Building the Docker image

While we can build the Docker image locally, this might take a long time given that we are reliant on hefty dependencies.

Using the cloud SDK, navigate to the folder of the Dockerfile. With the following snippet we can submit the Docker image to be built and stored in Google Cloud Container Registry.

export PROJECT_ID=$(gcloud config get-value project)
export IMAGE_URI=gcr.io/$PROJECT_ID/flair/text-classification:latest

# Build the container and submit to Cloud Container Registry.
gcloud builds submit --tag $IMAGE_URI

Store data in Cloud Storage

We can copy our local datasets to a Google Cloud Storage bucket by making use of the gsutil commands in the Cloud SDK.

For the sake of illustration we will copy .csv files to a bucket named flair-bucket and in the custom-container/dataset/ nested folder.

gsutil cp *.csv gs://flair-bucket/custom-container/dataset/

Testing the Docker

Rather than submitting the job immediately to AI platform, we should first test whether the Docker works. You can do this either locally, by using the Cloud SDK CLI or by making use of Cloud Shell.

Cloud Shell is the CLI in Google Cloud, and thus negates the fact that you would need to download the built image locally before you test it. Through this way we can quickly validate and debug, and saves us valuable time rather than having to wait on AI platform spinning up machines.

By navigating to https://console.cloud.google.com/ and run the following command in Cloud Shell.

export PROJECT_ID=$(gcloud config get-value project)
export IMAGE_URI=gcr.io/$PROJECT_ID/flair/text-classification:latest

# Run in Cloud Shell.
docker run $IMAGE_URI \
  --label_column_index 0 \
  --text_column_index 1   \
  --epochs 1   \
  --patience 1   \
  --gcs_data_path "custom-container/dataset/"  \
  --gcs_output_path "custom-container/output/" \
  --gcs_bucket_name "flair-bucket"   \
  --delimiter ","

Submit the training job

After we have verified that the Docker works. We can submit the container to AI platform.

For the full list of parameters, please refer to this documentation.

export PROJECT_ID=$(gcloud config get-value project)
export IMAGE_URI=gcr.io/$PROJECT_ID/flair/text-classification:latest
export REGION=europe-west1
export JOB_NAME=custom_container_job_$(date +%Y%m%d_%H%M%S)

# Submit training to AI platform using.
gcloud ai-platform jobs submit training $JOB_NAME \
  --region $REGION \
  --master-image-uri $IMAGE_URI \
  -- \
  --text_column_index 1   \
  --epochs 10   \
  --patience 3   \
  --gcs_data_path "custom-container/dataset/"  \
  --gcs_output_path "custom-container/output/" \
  --gcs_bucket_name "flair-bucket"   \
  --delimiter ","

Or if you make use of the Docker image which contains CUDA support, you can accelerate training by using a GPU.

In this example we attach a single NVIDIA Tesla K80 GPU by passing the BASIC_GPU to the scale-tier flag.

export PROJECT_ID=$(gcloud config get-value project)
export IMAGE_URI=gcr.io/$PROJECT_ID/flair/text-classification:latest
export REGION=europe-west1
export JOB_NAME=custom_container_job_$(date +%Y%m%d_%H%M%S)

gcloud ai-platform jobs submit training $JOB_NAME \
  --region $REGION \
  --master-image-uri $IMAGE_URI \
  --scale-tier BASIC_GPU \
  -- \
  --label_column_index 0 \
  --text_column_index 1 \
  --epochs 1   \
  --patience 1   \
  --gcs_data_path "custom-container/dataset/"  \
  --gcs_output_path "custom-container/output/" \
  --gcs_bucket_name "flair-bucket"   \
  --delimiter "," \
  --model_type "TransformerDocumentEmbeddings" \
  --model "sentence-transformers/xlm-r-100langs-bert-base-nli-stsb-mean-tokens"

Deployment

From my experience, deploying a fine-tuned text-classifier model based on sentence transformers on AI platform was quite a challenge:

In general, AI Platform requires you to build a deployment package in order to deploy models.
However, Pytorch models are not natively supported and are considered a custom prediction routine.
Given that you are making use of a custom prediction routine, the deployment package must be 500 MB or less. ** In fact, my best-model.pt file was already over 500 MB, excluding any required dependencies.
Finally, while there are machine types that support deployment packages up to 2 GB, they do not support custom prediction routines at this time.
In addition, these larger machine types do not scale down to 0, but will always have 1 node running.

As such, serving predictions on AI Platform was technically unfeasible and financially unattractive in comparison to Cloud Run

Cloud Run is a suitable alternative for this case is due to the following reasons:

By making use of Docker containers we can package our code in a straightforward way.
There is a seemingly generous free tier and you are billed per 100 milliseconds of uptime.
The service can scale down to 0.

Therefore, the example I present makes use of a Docker file to host the model on Cloud Run and uses FastAPI to serve predictions.

Overview

The Docker image installs the required libraries and creates the following folder structure:

/root/
   ├── app.py
   └── model
         └── best-model.pt

With app.py being the FastAPI endpoint that handled the incoming request, passes the text to the model, and returns the labels and probabilities.

Building the Docker image

While we can build the Docker image locally, this might take a long time given that we are reliant on hefty dependencies.

Using the cloud SDK, navigate to the folder of the Dockerfile. With the following snippet we can submit the Docker image to be built and stored in Google Cloud Container Registry.

export PROJECT_ID=$(gcloud config get-value project)
export IMAGE_URI=gcr.io/$PROJECT_ID/flair/text-classification-endpoint:latest

# Build the container and submit to Cloud Container Registry.
gcloud builds submit --tag $IMAGE_URI

Once Cloud Build is finished, you can create a Cloud Run endpoint with your container.

For the full list of parameters, please refer to this documentation.

export PROJECT_ID=$(gcloud config get-value project)
export IMAGE_URI=gcr.io/$PROJECT_ID/flair/text-classification-endpoint:latest
export REGION=europe-west1

# Deploy to cloud run.
gcloud run deploy text-classification \
          --image $IMAGE_URI \
          --allow-unauthenticated \
          --platform managed \
          --region $REGION \
          --timeout 60 \
          --memory 2Gi \
          --port 8080

Wait a few moments until the deployment is complete. After successful completion, the command line displays the service URL.

Testing the endpoint

After we have deployed the endpoint, we should test it before using it in production environments.

# Programmatically returns the endpoint url.
export ENDPOINT_URL=$(gcloud run services describe text-classification --platform managed --region europe-west1 --format 'value(
status.url)')

# Print the endpoint url
echo $ENDPOINT_URL

FastAPI also generates documentation of your endpoint at the /docs suffix of your url.

# Programmatically returns the endpoint url.
export ENDPOINT_URL=$(gcloud run services describe text-classification --platform managed --region europe-west1 --format 'value(
status.url)')

# Visit this url to check the automatically generated documentation of the endpoint.
echo $ENDPOINT_URL/docs

Let your newly created endpoint classify "Hello Flair."

# Programmatically returns the endpoint url.
export ENDPOINT_URL=$(gcloud run services describe text-classification --platform managed --region europe-west1 --format 'value(
status.url)')

# Make a post request to test the endpoint with "Hello Flair."
curl -X POST $ENDPOINT_URL/classify-text -H  "accept: application/json" -H  "Content-Type: application/json" -d "{\"text\":\"Hello Flair.\"}"

If everything is working correctly, the response of the model will be in the following form:

[{"_value":"label","_score":probability}]

Contributions & Suggestions

Pull requests and issues are welcome!

I am hoping to learn more and to improve the code.

Authors

Robin van Schaik

Acknowledgements

Flair for the text classification training framework.
Sentence transformers for great sentence-level language models that can be used in Flair.
Huggingface for a large collection of language models that can be used in Flair.

Awesome

flair-on-gcp

Table of contents

Getting started

Training

Overview

Building the Docker image

Store data in Cloud Storage

Testing the Docker

Submit the training job

Deployment

Overview

Building the Docker image

Testing the endpoint

Contributions & Suggestions

Authors

Acknowledgements