Great MLOps Tools to Launch Your Next Machine Learning Model
medium.com
medium.com
Not quite. We use MLFlow to allow our users to track their experiments, etc. We work around MLflow to spare users complexity. Saving a model can lead you to figure out that you need the exact same Python version for it to be deserialized, not even talking about the protocol, as MLflow uses "cloudpickle". It also relies on Conda, and it can make for a funny afternoon because installing Pytorch magically installs the latest version of Python without telling you. I had to wrap my head around that concept. That's like installing Thunderbird and suddenly it upgrades Ubuntu. That's not MLflow per se, but working with it users either have to deal with it, or we have to hide that complexity.
Deploying non trivial models expecting tensors doesn't work. By non trivial I mean a model that expects something else than a 2 dimensional DataFrame[0]. They're working on it, though.
This has never happened to me. I typically create a conda environment with a specific python version, then install Pytorch and it doesn’t update the python.