# download the runescape grand exchange CSVs
make download-data
# train a pytorch model
make train-pytorchMLflow now supports tagging models (giving mutable aliases) and is deprecating the old
way of using rigid stages (e.g. Production, Staging, Archived).
latest is a reserved tag. I'm not sure how you use it.
To fetch a model using the mlflow CLI. You can use
# fetch the item_1215 model of version 1
mlflow artifacts download \
--artifact-uri models:/item_1215/1 \
--dst-path ./modelYou can also assign the model an alias in the UI by running
# cd in to the directory where mlruns/ is located
cd ./example/
# run the ui; go to your browser and tag the model, e.g. with "challenger"
mlflow ui --port 5002
# download the model, this time using the aliased version
mlflow artifacts download \
--artifact-uri models:/item_1215@challenger \
--dst-path ./modelThe result of each of these is a file structure like this
./model
├── artifacts # model weights, reports, etc.
│ ├── model_config.json
│ ├── model_state.pt
│ ├── requirements-graph-full.txt
│ ├── requirements-graph-model.txt
│ ├── scaler.pt
│ └── uncommitted_changes.patch
├── code
│ └── train_pytorch # code needed to run inference
│ ├── __init__.py
│ ├── constants.py
│ ├── dataset.py
│ ├── envelope.py # load and use the model
│ ├── model.py
│ ├── py.typed
│ └── train.py
├── conda.yaml
├── input_example.json
├── MLmodel
├── python_env.yaml
├── python_model.pkl
├── registered_model_meta
├── requirements.txt # minimal reqs needed for inference
└── serving_input_example.jsonTODO
Track these aspects of each experimental run:
- Git metadata; capture: (note the solution requires the
gitexecutable to be available)- Git hash
- Git branch
- Remote URL
- Dirty/uncommitted changes (as patch)
- Log an HTML rendered diff of that
- Entrypoint
- the script: e.g.
uv run src/train_pytorch/train.py - CLI args, e.g.
--num-epocs 10
- the script: e.g.
- System metrics (docs)
- CPU
- Memory
- GPU
- Full python deps
- Python version
- Installed packages
- OS/non-Python deps
- Docker base image
Save this metadata with the model:
- Minimal 3rd Party Python deps
- Python version
- Installed packages
- Use pip-tools or uv to prune subtrees of dependencies, e.g.
jupyterlab
- Use pip-tools or uv to prune subtrees of dependencies, e.g.
- 1st Party Python code
- Artifacts
- Pyfunc wrapper
- Reference to the experiment run