This project is inspired by Tensorflow Recommenders. It is a demo of building a recommender model with a retrieval stage (not including ranking stage) that can provides recommendations of merchant based on the current session activities including the merchants a user have viewed and his search terms.
We use GCP as our infras for both training and serving.
- Notebooks that contain development in which experiments are tracked using Comet AI and model analysis to identify areas of improvement for later iterations
- Example model architecture is depicted in
model_architecture.md - Deployment of model artifact to GCP
- Pipeline written using Kubeflow to submit training job to GCP Vertex AI Pipelines. We have two kinds of pipeline here: one is for full retraining and one is for incremental retraining to save resources updating a model
- We use Vertex AI Serving for management of model endpoints after training
- The notebook
notebooks/experiment.ipynbhosts the common experiment workflow - The notebook
notebooks/pipeline.ipynbpackages the training code and compile a scheduled pipeline
Please reach out to dvquy.13@gmail.com if you want to discuss anything about this repo.