dfml

sklearn-style ML preprocessing with pandas-native syntax.

dfml registers a .ml accessor on every pandas DataFrame — the same mechanism pandas itself uses for .str, .dt, and .cat — so common preprocessing steps read like native pandas instead of sklearn boilerplate.

import pandas as pd
import dfml  # noqa: F401  (side-effect import registers the accessor)

df = pd.read_csv("data.csv")
clean = df.ml.impute().ml.encode().ml.scale()

Why

Preprocessing with plain sklearn usually looks like this:

numeric_features = X.select_dtypes(include=["int64", "float64"]).columns
categorical_features = X.select_dtypes(exclude=["int64", "float64"]).columns

numeric_transformer = Pipeline([
    ("imputer", SimpleImputer()),
    ("scaler", StandardScaler()),
])
categorical_transformer = Pipeline([
    ("imputer", SimpleImputer(strategy="most_frequent")),
    ("onehot", OneHotEncoder()),
])

preprocessor = ColumnTransformer([
    ("num", numeric_transformer, numeric_features),
    ("cat", categorical_transformer, categorical_features),
])

X_transformed = preprocessor.fit_transform(X)  # numpy array, column names gone

With dfml:

clean = df.ml.impute().ml.encode().ml.scale()  # still a DataFrame, columns intact

• Stays a DataFrame. sklearn transformers return bare numpy arrays and drop your column names. Every dfml method returns a DataFrame.
• Chainable. .ml.impute().ml.encode().ml.scale() reads top to bottom like a pipeline, without building one.
• Auto-detects columns. Numeric vs. categorical columns are inferred by default — pass columns=[...] only when you want to override it.
• Not a replacement. Every method wraps a real sklearn transformer under the hood (accessible via .ml.fitted_), so you can drop back to raw sklearn any time you need something the accessor doesn't cover.

Install

pip install dfml

Usage

Impute missing values

df.ml.impute(strategy="mean")               # numeric columns, default
df.ml.impute(strategy="most_frequent")       # all columns
df.ml.impute(columns=["age"], strategy="median")

Encode categorical columns

df.ml.encode()                               # one-hot, all non-numeric columns
df.ml.encode(columns=["city"], method="label")
df.ml.encode(columns=["city"], drop_first=True)

Scale numeric columns

df.ml.scale()                                # StandardScaler, all numeric columns
df.ml.scale(method="minmax")
df.ml.scale(columns=["income"], method="robust")

Split into train/test

X_train, X_test, y_train, y_test = df.ml.split(target="churn", test_size=0.2)

Chain everything

X_train, X_test, y_train, y_test = (
    df.ml.impute()
      .ml.encode()
      .ml.scale()
      .ml.split(target="churn", test_size=0.2)
)

Inspect what was fitted

scaled = df.ml.scale()
scaled.ml.fitted_["scale"]["transformer"]   # the actual fitted StandardScaler

Development

git clone https://github.com/Parthshewale18/dfml
cd dfml
pip install -e ".[dev]"
pytest

License

MIT