Configuration file

Training options live in a TrainConfig, which can be written as YAML or JSON and passed with amesml train --config FILE or TrainConfig.from_file(FILE). Unknown keys are rejected, so a typo cannot silently fall back to a default.

# Every TrainConfig option, with its default. Use with:
#   amesml train --config examples/config.example.yaml
# Command-line --task and --model override the values here.

task: regression            # regression | classification
model: gbm                  # linear | ridge | lasso | gbm
target: SalePrice
id_column: Id
test_size: 0.2
random_state: 42

# Classification only: a price, or "median" for the training rows' median.
threshold: median
stratify: false             # classification only

# Derived features: total area, house and remodel age, bathrooms, porch area.
engineer_features: true

# Leave unset to infer column types from the data.
numeric_features: null
categorical_features: null
categorical_codes: [MSSubClass, MoSold]
zero_fill_features: [GarageYrBlt, MasVnrArea]

drop_first: true
alpha: null                 # ridge/lasso strength; null = cross-validated default
max_iter: 5000

Options

Option	Default	Meaning
task	regression	regression predicts the price; classification predicts at/above threshold
model	linear	linear, ridge, lasso or gbm (see below)
target	SalePrice	target column
id_column	Id	identifier, never used as a feature
test_size	0.2	share of rows held out by train
random_state	42	seed for the split and every seeded estimator
threshold	median	classification: a price, or the training rows' median
stratify	false	classification: keep the class balance in both splits
engineer_features	false	add total area, ages, bathrooms and porch area
numeric_features	inferred	use exactly these numeric columns
categorical_features	inferred	use exactly these categorical columns
categorical_codes	[MSSubClass, MoSold]	numeric columns that are category codes
zero_fill_features	[GarageYrBlt, MasVnrArea]	numeric columns where missing means "none"
drop_first	true	drop one level per one-hot encoded column
alpha	per model	regularisation strength for ridge and lasso
max_iter	5000	iteration limit for iterative solvers

The defaults reproduce the original notebook closely - to the cent for the classification, and to within about 5% for the regression figures, which come from an iterative solver whose stopping point shifts with the scikit-learn, scipy and BLAS build.

Model families

model	Regression	Classification	Default alpha
linear	ordinary least squares	logistic regression (C = 1)	-
ridge	ridge regression	L2-penalised logistic regression	10 / 30
lasso	lasso	L1-penalised logistic regression	300 / 3
gbm	histogram gradient boosting	histogram gradient boosting	-

For classification C = 1 / alpha, so in both tasks a larger alpha means a simpler model. The defaults were chosen by 5-fold cross-validation on the training split; see Design decisions.

Overriding feature types

Setting numeric_features or categorical_features replaces the inferred list for that role only; the other role is still inferred from the remaining columns. To train on a handful of columns, set both:

numeric_features: [GrLivArea, OverallQual, YearBuilt, TotalBsmtSF, GarageCars]
categorical_features: [Neighborhood]

With engineer_features: true the engineered columns are added to the numeric features, and the raw columns they need are read even if not listed.