scikit-learn

FloDR is a scikit-learn transformer. It clones, takes part in pipelines and model selection, and returns DataFrames on request.

Pipelines

from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

pipe = make_pipeline(StandardScaler(), FloDR(random_state=0)).fit(X)
Y_new = pipe.transform(X_new)

Only scale the input when that is the geometry you want. See Choosing the input.

DataFrames

model = FloDR(random_state=0).set_output(transform="pandas").fit(df)
model.transform(df)                 # DataFrame with columns flodr0, flodr1
model.feature_names_in_             # the column names of df
model.get_feature_names_out()       # ["flodr0", "flodr1"]

Model selection

score returns the mean log density, so GridSearchCV and cross_val_score select on held-out likelihood when no scorer is given. See Density.

Pickling

A fitted model pickles with its tensors on the CPU. A model fitted on a GPU therefore loads on a machine without one and moves back to the GPU where one is available.

import pickle

pickle.dump(model, open("model.pkl", "wb"))
model = pickle.load(open("model.pkl", "rb"))

Estimator checks

FloDR passes scikit-learn's check_estimator apart from two groups of checks. Several checks force n_components=1, and a one-dimensional layout leaves the flow no residual to route through. Two checks expect transform on the training data to reproduce fit_transform, which does not hold for padded inputs. The test suite lists both groups as expected failures.

Deprecated arguments

fit(X, iters=...), fit(X, progress=...) and inverse_coords still work and warn. Use max_iter, verbose and inverse_latent instead.