Let's Dockerize TPOT!
Imagine a quick and dirty SageMaker Autopilot. On your local. In Docker.
The following implements the "Digits dataset example" (i.e., MINST) in the TPOT documentation.
I don't know the TPOT API well at all, so any errors of commission or omission are due to my inexperience with it.
Only the usage with make has been documented below.
Take a look at the Makefile to see the commands corresponding to the targets.
- Build the image.
make build
- To see how to run a Jupyter notebook in the container run this.
make jupyter
After you do, you should see something like the following at the command line.
docker run --publish 8888:8888 --volume $(pwd)/notebooks:/workspace/notebooks --detach dockerpot jupyter
68634eea7fcb08ddf751e162ff91f0eb651c6c15ae8d2763e6b5ebb7d06e49b8
-
Go to http://localhost:8888/notebooks/notebooks/run_minst.ipynb, and run the notebook.
-
Find tpot_exported_pipeline.py in the notebooks directory.
-
Stop the container.
make stop
- Run the run_minst.py script.
make run_minst
After you do, you should see something like this at the command line.
docker run --volume $(pwd)/scripts:/workspace/scripts dockerpot python /workspace/scripts/run_minst.py
Optimization Progress: 2%|▏ | 2/120 [00:01<01:36, 1.22pipe
When it's done, you'll see something like this.
docker run --volume $(pwd)/scripts:/workspace/scripts dockerpot python /workspace/scripts/run_minst.py
Generation 1 - Current best internal CV score: 0.9591656340355226
Generation 2 - Current best internal CV score: 0.9658405617513424
Generation 3 - Current best internal CV score: 0.9710422690348341
Generation 4 - Current best internal CV score: 0.9866363761531047
Generation 5 - Current best internal CV score: 0.9888641057414291
Best pipeline: KNeighborsClassifier(XGBClassifier(input_matrix, learning_rate=1.0, max_depth=9, min_child_weight=11, n_estimators=100, nthread=1, subsample=0.7500000000000001), n_neighbors=1, p=2, weights=distance)
0.9844444444444445
- Find tpot_exported_pipeline.py in the scripts directory.
Use make docker-clean-unused to nuke containers on your machine that aren't running. Watch out.
Use make docker-clean-all to kill all containers, running or not. Careful.
Use make shell to peek inside the container.