Skip to content

Parametrized experiments#

Fixed and variable parameters#

Parametrized experiments allow you to control and differentiate the behaviour of runs:

  • Parameter grids are grids of parameters with a discrete number of values for each, can be set on runs using Experiment.add_runs(A=[1,2,...], B=...) and are accessible within the run at run.params.

  • Configuration parameters are fixed for all runs, can be accessed at run.config and are set by passing config={A=1, B=2, ...} to Experiment.execute. The value is the same for all runs.

In the next example:

  1. We define a fixed configuration parameter X. We use a parameter grid, similarly to sklearn.model_selection.ParameterGrid, to define run parameters A and B.

  2. We execute the step function step_sum, summing up the value of A, B and X.

  3. We persist and query the experiment as a Pandas dataframe.

Creating an experiment

from mltraq import Run, create_experiment

def step_sum(run: Run):
    run.fields.sum_ABX = run.params.A + run.params.B + run.config.X

experiment = (
    .add_runs(A=[1, 10], B=[100, 1000])
    .execute([step_sum], config={"X": 10000})

                                 id_run  sum_ABX
0  d65df69e-1175-44a5-be2f-2232765703b9    10110
1  d65df69e-1175-44a5-be2f-2232765703ba    11010
2  d65df69e-1175-44a5-be2f-2232765703bb    11001
3  d65df69e-1175-44a5-be2f-2232765703bc    10101


The attributes run.config and run.params are not persisted to database by default. You can request its transparent persistence with Experiment.execute(...., args_field="args"), which will persist/reload config/params to/from run.fields.args.


You know how to define parameter grids and explore the impact of varying parameters on evaluation metrics.