In the 3rd session, we will harness the pipeline concept towards manageable high throughput experimentation in ML/DL research.
Currently, complex pipelines are found in the field of ML in implementations of automated training and deployment of ML models. However, these pipelines and the code they encapsulate are rarely those that are used in the research stage. Moreover, existing research pipelines tend to be focused on the data preparation stage, and are mostly trivial afterward.
This represents several areas where we can do better:
* Easily “grow” automated multi-stage workflows from research code with minimal code changes.
* Frictionless executions of these pipelines on available resources.
* Minimizing re-writes when promoting code from research towards “production”
I will address suggestions for improvements in these, with specific examples from simple to intricate workflows in research.
In the previous webinar, we established how ensuring reproducibility in ML research enables automation, which in turn unlocks advanced MLOps (such as pipelines). Since these topics will be used for this webinar, it is recommended to refresh your memory with the recorded event.
All sessions of the series: