ANIME RECOMMENDATION APPLICATION

Project Organization

├── docker             <- folder with all the useful Dockerfiles
│
├── flask              <- Flask API used to expose the model
│
├── models             <- Trained and serialized models
│
├── src                <- Source code for use in this project.
│
├── streamlit          <- Streamlit user interface 
│
├── Makefile           <- Makefile with commands like `make train`
│
├── README.md          <- The top-level README for developers using this project.
│
├── requirements.txt   <- The requirements file
│
├── setup.py           <- makes project pip installable (pip install -e .) so src can be imported

---

Project based on the cookiecutter data science project template. #cookiecutterdatascience

How to run the project

1. Environment Initialization

1.1. Python requirements

pip install -r requirements.txt

1.2. Environment Variables

The variables used in this README are all defined in the .env file.
These variables are automatically imported in the Makefile but if in the following parts you want to run any of the command manually you have to run either of these command lines

make init_env_variables

or with the full command

export $(grep -v '^#' .env | xargs)

In both cases, you also need to store your local IP adress.

LOCAL_IP_ADRESS=$(ipconfig getifaddr en0)

This needs to be run every time you start a new terminal.

2. Get the data

The data used in this project comes from a Kaggle dataset Anime Recommendation Database 2020.
For the training of this app you will need anime.csv and rating_complete.csv to be stored in the ./data/ folder.

3. Train the model

make train

or with the full command

python src/training.py \
    --data_folder $(PWD)/data \
    --model_path $(PWD)/models/trained_model.pkl \
    --nb_hidden_features $NB_FEATURES

This will take a couple of minutes depending on your own computer configuration but also on the number of features used.

4. Deploy the model

4.1 Launch the Flask API

make flask_run

or with the full command

python flask/app.py \
    --model_path=$(PWD)/models/trained_model.pkl \
    --debug=False \
    --host_ip=$LOCAL_IP_ADRESS \
    --port=$FLASK_PORT

The server should be ready to use in a few seconds.

4.2 Test it (optional)

make predict_shingeki_no_kyojin

or with the full command

curl -X POST http://$LOCAL_IP_ADRESS:$FLASK_PORT/predict \
    -d @flask/samples/shingeki_no_kyojin.json \
    -H "Content-Type: application/json"

With this you should get a recommendation based on Shingeki no Kyojin (Attack on Titans) ratings.

5. Deploy the user interface

5.1 Launch the streamlit app

make streamlit_run

or with the full command

streamlit run streamlit/app.py \
    --browser.serverAddress $LOCAL_IP_ADRESS \
    --server.port $STREAMLIT_PORT \
    -- \
        --anime_path $(PWD)/data/anime.csv \
        --model_ip $LOCAL_IP_ADRESS \
        --model_port $FLASK_PORT

5.2 Try it out!

If no tab started on your brower you can run this command and paste the result in your favorite browser.

echo http://$STREAMLIT_IP:$STREAMLIT_PORT

6. Running previous services with Docker

6.0 Install Docker

Install docker here

6.1 Training

6.1.1 Build

make docker_build_training

or with the full command

docker build \
    -t training:v0 \
    -f docker/training/Dockerfile .

6.1.2 Run

make docker_run_training

or with the full command

docker run \
    -it --rm \
    -v $(PWD)/data:/anime/data \
    -v $(PWD)/models:/anime/models \
    training:v0 \
        --data_folder /anime/data \
        --model_path /models/trained_model.pkl \
        --nb_hidden_features $NB_FEATURES

Note: I am having huge performance issues when running the training on a docker so unless you run it on the cloud you may have to train the model without docker (see 3.)

6.2 Flask

6.2.1 Build

make docker_build_flask

or with the full command

docker build \
    -t flask:v0 \
    -f docker/flask/Dockerfile .

6.2.2 Run

make docker_run_flask

or with the full command

docker run \
    -p $FLASK_PORT:$FLASK_PORT \
    -v $(PWD)/models:/anime/models \
    flask:v0 \
        --model_path=./models/trained_model.pkl \
        --debug False \
        --host_ip $LOCAL_IP_ADRESS \
        --port $FLASK_PORT

Note: Same as for the training, I am having performance issues when running the training on a docker locally so I would advise you to run flask without docker (see 4.).

6.3 Streamlit

6.2.1 Build

make docker_build_streamlit

or with the full command

docker build \
    -t streamlit:v0 \
    -f docker/streamlit/Dockerfile .

6.2.2 Run

make docker_run_streamlit

or with the full command

docker run \
    -p $STREAMLIT_PORT:$STREAMLIT_PORT \
    -v $(PWD)/data:/anime/data \
    streamlit:v0 \
        --browser.serverAddress $LOCAL_IP_ADRESS \
        --server.port $STREAMLIT_PORT \
        -- \
            --anime_path ./data/anime.csv \
            --model_ip $LOCAL_IP_ADRESS \
            --model_port $FLASK_PORT

For the streamlit app the performances are not harmed at all by the use of docker.

To improve

If I were to continue this project, I would:

• Find a way to make the docker containers more efficient
• Add command lines to train and expose the model/app on the cloud
• Test Poetry to handle requirements
• Bonus: Refacto the code