Experiments with ML Flow

Experiment 1: Getting started

• Install: pip3 install mlflow jupyter torch torchvision matplotlib
• Open jupyter with jupyter-notebook
• Try to run the notebook.
• Open web UI with mlflow ui
• You should see your test_run1 on the list of runs.
  • Run is either until the end of the Python process or when the context of mlflow.start_run(...) ends.

Experiment 2: Remote tracking server setup

• Build the docker image: cd Experiment2 && docker build -t mlflow-experiment:latest .
• Run the remote tracking server: docker run -p 5000:5000 mlflow-experiment:latest
• Start the jupyter notebooks and open Experiment2.ipynb

Experiment 3: Write a model in PyTorch and track it's training in MLFlow

• Run the remote tracking server: docker run -p 5000:5000 mlflow-experiment:latest
• Start the jupyter notebooks and open Experiment3.ipynb

Experiment 4: Serialize the model to the MLFlow format

• In experiment 3 we saved a model in cifar10_model/cifar10.mlflow directory
• This model can be served via this command: mlflow models serve --no-conda -m cifar10.mlflow -p 5001
• The API is served the at http://localhost:5001/invocations
  • e.g. you can call it with (mind that input format is not aligned to our model in this example) curl http://localhost:5001/invocations -H 'Content-Type: application/json' -d '{"columns": ["a", "b", "c"], "data": [[1, 2, 3], [4, 5, 6]]}'

Screenshots