Quickstart

This fits FloDR to a matrix with one row per point, then uses each part of the fitted model once.

import numpy as np
from flodr import FloDR, viz

rng = np.random.default_rng(0)
X = rng.normal(size=(5000, 50)).astype("float32")
labels = (X[:, 0] > 0).astype(int)

model = FloDR(random_state=0).fit(X)
Y = model.embedding_                        # (5000, 2) layout

Y_new = model.transform(X[:100])            # one pass through the flow
Y_opt = model.transform_opt(X[:100])        # placement refined against the training layout

Z = model.transform_latent(X[:100])         # layout followed by the residual
X_back = model.inverse_latent(Z)            # equal to X[:100] up to float precision
X_rep = model.inverse_transform(Y[:100])    # one representative input per position

logp = model.score_samples(X[:100])         # log density of each point

sigma = model.conditional_spread()          # spread of the inputs behind each position
cert = model.spread_calibration()           # held-out test of that field
field, hidden = model.hidden_contrast(labels)

ax = viz.plot_embedding(Y, labels)
viz.save(ax.figure, "embedding.png")

A fit of 5,000 points takes about two minutes on a consumer GPU. hidden_contrast trains a classifier for every permutation of its null and is the slowest call here.

The layout depends on the input more than on any parameter, so read Choosing the input before fitting your own data. Fitting explains the parameters, and Diagnostics explains how to read the fields and their tests.

Next​

Next