A mini project that stages, models, generates, and classifies Apple SmartWatch Data. We will use AWS S3, AWS EMR, Kafka, Apache Spark, and Databricks Datagen.
The first stage of the application is to load bulk data (~160GB) from Washington State University (WSU)
Center of Advanced Studies in Adaptive Systems (CASAS) archives
https://casas.wsu.edu/datasets/smartwatch/ into
a staging folder in an S3 bucket. We use a Spark job on EMR to clean the data (remove null values) and save the processed
data into a processed folder on S3 and also into a glue table.
The cleaned up data is used to train a machine learning classifier that predicts the recorded activity. The spark script that performs this can be found here Activity Classifier The trained model will be used to predict the activity of new generated samples streamed from a kafka topic using spark streaming later in the application.
To retrain the model with new samples, active learning technique is used. When new data samples are streamed from kafka topic, the activities on those samples are predicted using the model trained above. Misclassified samples are labeled with the correct class labels by a human annotator and saved to a feature store. Airflow routinely picks samples in the feature store and retrains the model. If the new model has a better performance than the old one, it is saved and used to classify new streaming samples
To simulate the generation of new Apple Watch Data, we will use Databricks Labs DataGen and the descriptive statistics calculated from the cleaned data samples. This ensures that generated data maintains the distribution of the original dataset. The generated data is loaded into spark streaming and written into a kafka topic. The activity prediction pipeline above picks up the data from the kafka topic and makes predictions. The samples with low predictions are sent to the annotation pipeline that will be used to retrain the model at a later date. Predictions are saved into a SQL database (like Redshift)
To run, clone the project. Run the init.sh to fetch data from WSU and sync it to S3. The script also copies spark scripts
from local directory to s3 bucket to be used later in EMR, and builds the project using docker compose build.
Run rundatapreprocess.py to submit an EMR job which creates a cluster (if it doesn't exist), and cleans up the watch
data in S3. The cleaned data is saved into a processed folder on S3, and a Glue table that will later be migrated to
a redshift table.
python activitymodule/jobs/run-data-preprocess.py
This job cleans up data from the staging directory on S3 and saves the result as parquet file on S3. It also saves a copy to a glue table (and later redshift).
Run train-activity-classifier.py to submit the EMR to train a machine learning classifier to predict watch activity
using the cleaned data.
Start the project using docker compose up. The spark container runs the start-jobs.sh script which submits the
predict-activity.py and generate-new-data.py using spark-submit
/usr/lib/spark/bin/spark-submit /home/hadoop/spark/scripts/predict-activity.py
/usr/lib/spark/bin/spark-submit /home/hadoop/spark/scripts/generate-new-data.py
The generate-new-data script uses dblDataGen to generate new watch data that is of the same distribution as the previously
loaded watch data. This statistics of the data is loaded from glue or redshift. The newly generated data is sent into a
kafka topic.
The predict-activity script reads data from kafka topic and predicts the activity associated with the data.
The machine learning model trained above is loaded into the predict-activity notebook which also starts a spark
streaming job. The spark stream takes data from ticker-watch-data kafka topic which is constantly loaded with data
generated by the generate-new-data script above
View the kafka topics, spark scripts, and cluster performance using the following routes:
© | @deramos