deepsim-dskit - A Reusable Data Science Framework

deepsim-dskit is an installable Python package for reproducible, configuration-driven data science pipelines. It provides reusable building blocks for loading data, preprocessing, splitting, modeling, artifact management, and experiment runs.

Installation

pip install -e ".[dev]"

Optional extras:

pip install "deepsim-dskit[polars]"
pip install "deepsim-dskit[yaml]"

Quick Start

from dskit import load_dataset, create_split

df = load_dataset("data/advertising.csv", index_col=0)
split = create_split(df, target="sales", test_size=0.2, random_state=42)

Run a full experiment from a config dictionary:

from dskit import run_full_pipeline

config = {
    "experiment_id": "advertising_baseline",
    "seed": 42,
    "data": {
        "path": "data/advertising.csv",
        "target": "sales",
        "read_kwargs": {"index_col": 0},
    },
    "splitting": {"test_size": 0.2, "val_size": 0.1, "random_state": 42},
    "preprocessing": {
        "missing": {"strategies": {}, "indicator_columns": []},
        "outliers": {"columns": [], "method": "iqr", "multiplier": 1.5},
        "scaling": {"columns": ["TV", "radio", "newspaper"], "method": "standard"},
    },
    "models": {
        "linear": {"class": "LinearRegression", "params": {}},
        "ridge": {"class": "Ridge", "params": {"alpha": 1.0}},
    },
    "output": {
        "experiments_dir": "experiments",
        "registry_path": "registry/experiments.json",
    },
}

result = run_full_pipeline(config)
print(result["best_model_name"])
print(result["metrics"]["test_r2"])

The output block may also include logs_dir (defaults to "logs").

CLI

dskit-run --version
dskit-run --config configs/advertising.json --dry-run
dskit-run --config configs/advertising.json --env production

What's Included

License

MIT License. See LICENSE.

Author

Shouke Wei, PhD · Deepsim Press Author Page Affiliation: Deepsim Intelligence Technology Inc. deepsim.ca