A new model of your own

For starting a new forecasting model: chap-devops models new makes the project, Docker builds its image, and a chap-devops deployment trains it, predicts with it and scores it through chap-core. Python, uv, R and the chapkit CLI are not necessary on the machine.

Make the project

chap-devops models new dengue-lags

created dengue-lags in dengue-lags (fn-py template, service id dengue-lags)
next: build the image as `dengue-lags/README.md` says, then run `chap-devops run dengue-lags:dev`

The default type, fn-py, is a chapkit service with the model in Python, in main.py. For a model in R scripts, add --template shell-r. For a model in the MLproject format, use --template mlproject-py or --template mlproject-r. The types lists all nine.

The example model predicts, for each location, the mean of disease_cases in the last window periods. Replace it with your model. README.md in the project lists the files to change.

Build the image

In the project:

cd dengue-lags
git init && git add . && git commit -m "feat: first version"
docker build --platform linux/amd64 --build-arg GIT_REVISION=$(git rev-parse HEAD) -t dengue-lags:dev .
cd ..

chap-core stores a model only if the image has a git revision, so build from a commit with GIT_REVISION. The first build pulls the base image, ghcr.io/dhis2-chap/chapkit-py. A rebuild takes some seconds.

Run it with chap-core

chap-devops init lab --models none
cd lab
chap-devops models add dengue-lags:dev
chap-devops up --wait
chap-devops status

chap-devops up --wait returns when the model has registered, and it creates the configured model default. chap-devops status then shows:

chap-core   up   http://localhost:8700   2.4.0   auth: off

MODEL        STATE       REACH          LAST PING
dengue-lags  registered  via chap-core  3s ago

1 model registered

Test it through chap-core

chap-devops models test dengue_lags --backtest --seed 7

testing 1 model (through chap-core: a dataset, a backtest and its scores)
dengue-lags    pass      16s   crps 35.5  mae 35.5  rmse 38.8
  configured model: default (id 14)

1 of 1 model passes

The test makes a dataset from the model’s own sample data, runs a backtest in chap-core, and prints its scores. --seed makes the data the same on every run, so two runs can be compared.

Give an option another value

window is an option of the model, with 3 as the default. A configured model gives it another value:

chap-devops models configs add dengue_lags --name long --set window=6
chap-devops models test dengue_lags --backtest --seed 7 --config long

created configured model long of dengue_lags
testing 1 model (through chap-core: a dataset, a backtest and its scores)
dengue-lags    pass      13s   crps 41.4  mae 41.4  rmse 43.2
  configured model: long (id 15)

1 of 1 model passes

The Modeling App shows long as a model to pick, next to default. See Configured models.

After a change

chap-core stores each version of a model only once. If you build a new commit under the same version, the model registers, but its backtests fail:

dengue-lags    FAIL       8s   ValueError: Model template 'dengue-lags' version '0.1.0' is stored from revision 'db139b6b09185e0be0971d31cafb39f126c910...

So, before you build again, raise version in main.py (in MLproject for an MLproject type), for example to 0.1.1. Then commit, and build with the same docker build line. In lab:

chap-devops restart
chap-devops models test dengue_lags --backtest --seed 7

dengue-lags    pass      13s   crps 39.5  mae 39.5  rmse 42.3
  configured model: default (id 16)

created configured models in chap-core for dengue_lags (default)
1 of 1 model passes

In this run, the change was window: int = 4 in main.py, so the score is different. The backtest created the configured model default of version 0.1.1. A configured model belongs to one version, so long is not there for 0.1.1: add it again with chap-devops models configs add.

To run the model service alone, with no deployment: chap-devops run dengue-lags:dev. To publish the image, push the project to GitHub: the workflow in .github/workflows/publish.yml builds it and publishes it to GHCR. Then chap-devops models add takes the repository URL, as in Models outside the marketplace.

These lines come from chap-devops 0.101.0 with this change, chap-core 2.4.0 and chapkit 2.3.1 on macOS (arm64, the image under amd64 emulation), on 2026-10-08.

More: A new model project.

All shapes: Use cases.

Keyboard shortcuts