After completing this repository, you will be able to understand the following concepts:

• Provision an AWS EC2 and running the training of a BERT model with CML
• Apply DagsHub Direct Data Access to improve your training process.
• Implement a Github actions pipeline using the previous instance.
• Automatically log your models metrics with MLFlow.
• Compare the model performance with MLflow Experiment.
• Automatically save your training metadata on DVC for easy tracking.

This approach can take significant amount of time to get the data from pull instruction.

Streaming simplifies the complexity of data collection by introducing a parallel computation approach, which starts the model training while getting the data from the storage.

A free tier is enough for this use case. Below are the properties of the EC2 instance used for the use case.

• cloud type: t2.micro.
• cloud-region: us-east-1a

• Access to DagsHub
• Access to GitHub Actions

• Python 3.9.1
• DVC 2.11
• CML
• You can find all the additional information in the requirements.txt file

DagsHub provides the capabilities to use MLFlow and DVC while giving the choice of working on Github. The following results are the experiments from DagsHub, using MLFlow to track the model F1-Score, Precision and Recall.

The following graphics shows the performances of the models for 3 epochs.

Below is the general workflow of the pipeline.