Sweeps: Hyperparameter Optimization Engine

This repo contains the routines that generate hyperparameter sweep suggestions in the W&B backend and client local controller.

Issues are not enabled in this repository. Please open issues related to sweeps in the wandb client library github issues page.

Installation

To install:

pip install sweeps

Examples

Get next run in a sweep.

Requires two arguments, config, the config that defines the sweep, and runs, the other runs in the sweep

config:

{
    "metric": {"name": "loss", "goal": "minimize"},
    "method": "bayes",
    "parameters": {
        "v1": {"min": 1, "max": 10},
        "v2": {"min": 1.0, "max": 10.0},
    },
}

runs:

[
    SweepRun(
        name="b",
        state=RunState.finished,
        history=[
            {"loss": 5.0},
        ],
        config={"v1": {"value": 7}, "v2": {"value": 6}},
        summary_metrics={"zloss": 1.2},
    ),
    SweepRun(
        name="b2",
        state=RunState.finished,
        config={"v1": {"value": 1}, "v2": {"value": 8}},
        summary_metrics={"loss": 52.0},
        history=[],
    )
]

Codepath:

suggestion = next_run(config, runs)

next_run:

validates that sweep config conforms to the jsonschema in config/schema.json, if not, it raises a ValidationError
parses the config file and determines the method that it should use to find the next run (in this case bayes_search_next_run)
calls bayes_search_next_run(config, runs) and returns the suggested SweepRun

Return list of runs to stop in a sweep.

Requires two arguments, config, the config that defines the sweep, and runs, the other runs in the sweep

config:

{
    "method": "grid",
    "metric": {"name": "loss", "goal": "minimize"},
    "early_terminate": {
        "type": "hyperband",
        "max_iter": 5,
        "eta": 2,
        "s": 2,
    },
    "parameters": {"a": {"values": [1, 2, 3]}},
}

runs:

[
    SweepRun(
        name="a",
        state=RunState.finished,  # This is already stopped
        history=[
            {"loss": 10},
            {"loss": 9},
        ],
    ),
    SweepRun(
        name="b",
        state=RunState.running,  # This should be stopped
        history=[
            {"loss": 10},
            {"loss": 10},
        ],
    ),
    SweepRun(
        name="c",
        state=RunState.running,  # This passes band 1 but not band 2
        history=[
            {"loss": 10},
            {"loss": 8},
            {"loss": 8},
        ],
    ),
    SweepRun(
        name="d",
        state=RunState.running,
        history=[
            {"loss": 10},
            {"loss": 7},
            {"loss": 7},
        ],
    ),
    SweepRun(
        name="e",
        state=RunState.finished,
        history=[
            {"loss": 10},
            {"loss": 6},
            {"loss": 6},
        ],
    ),
]

Codepath:

to_stop = stop_runs(config, runs)

stop_runs:

validates that sweep config conforms to the jsonschema in config/schema.json, if not, it raises a ValidationError
parses the config file and determines the method that it should use to early terminate runs (in this case hyperband_stop_runs)
calls hyperband_stop_runs(config, runs) and returns the SweepRuns to stop