Home

Awesome

Tensorflow-Project-Demo

image For this project, I used the classic datasets iris to do the project demo. My personal goal for this project is to not only explore the data and build models, but to also build an API server with retrainable model. To achieve this goal, I used fastapi, tensorflow and ray tune.

The project contain three services db, serving and training. When you train the model. the training module will query the db to get the iris data and start training the model. After the training process finish, it will give the model to serving. So that any client service can connect to serving to get the prediction with given features.

Also, I decided to develop this project to be the same as how data-related projects are developed in real-world scenarios, wherein the end goal of development is a project that is feasible for production. Therefore, I have put efforts on creating:

  1. Exploratory Data Analysis (EDA) in the notebooks/ folder;
  2. An API Server inside the api/ folder;
  3. Files for deployment such as Dockerfile and docker-compose.yml;
  4. Documentations in the docs/ folder; and
  5. Some necessary scripts in scripts/ folder; and
  6. Configurations, included hyperparameters space in config.py.

In the tensorflow part, I used very powerful tools:

  1. tfrecord
  2. tensorboard
  3. tensorflow serving

for building the project.

Services

I have three services: db, serving and training.

Database

I use the PostgreSQL as this demo's database. After you run all the service Training will initial the database, then you can check the data in the database.

docker-compose exec db psql -d database -U user
select * from iris;

image

Training

image The training service is a machine learning API that is open on port 8000. I used fastapi for the API server, so you can check it on http://localhost:8000/docs after you run the training service. The hyperparameters search space is in the src/config.py so you can change them if you want.

The Retraining process start from query the db and then it will write the data on local as tfrecord format. This move is able to reduce lots of memory usage during training. And tfrecord is a binary file therefore in normal case, it can also reduce the data storage problem.

Serving

I use Tensorflow Serving to build the prediction server. It is a very powerful tools amount TensorFlow Extended (TFX) . See more detail on tensorflow document.

How to run this demo

1. Install requirement

2. Build the image

make build

3. Start the services

docker-compose up -d

4. Try to update the model

curl -X PUT "http://localhost:8000/model" -H  "accept: application/json"

5. Check Training Result

5.1 Activate tensorboard

docker-compose exec training make activate-tensorboard

5.2 Check the Result

Turn on your browser and go to http://localhost:6006/ image

6. Let's make some prediction

Feel free to change the number from 0 to any number you want

curl -X POST "http://localhost:8501/v1/models/tensorflow-project-demo:predict" 
-H  "accept: applicationjson"
-d 
'{
    "inputs":{
        "sepal_length":[[0]],
        "sepal_width":[[0]],
        "petal_width":[[0]],
        "petal_length":[[0]]
    }
}'

The response will be like:

{
    "outputs": [
        [
            0.607119262,
            0.392192692,
            0.000688050292
        ]
    ]
}

The output means the estimation in probability of categories Setosa, Versicolor and Virginica respectively.

Note

If you want to change the model structure, you can change the model.py. Also, If you want to change the hyperparameters search space, you can go to config.py.

About the hyperparameter search algorithm, I am using the random search for this demo but if you want to try other searching algorithm, you can change train.py.

I have another project use lightgbm as the back bone model. Take a look about the project lightgbm-project-demo

Here are some documentations

How to set up the working environment for this project

API example

Reference