Inference

fastcxt inference is a single forward pass per pair – no autoregressive sampling, no stochastic averaging.

From a tree sequence

from fastcxt.translate import translate_from_ts

means, variances, index_map = translate_from_ts(
    ts, model,
    pivot_pairs=[(0, 1), (2, 3), (0, 4)],
    mutation_rate=1e-8,
    device="cuda:0",
    batch_size=256,
)

# means: (N, W) log-TMRCA predictions
# variances: (N, W) prediction variances
# index_map: (N, 2) mapping to [block_idx, pair_idx]

From a genotype matrix

from fastcxt.translate import translate_from_genotype_matrix

means, variances, index_map = translate_from_genotype_matrix(
    gm=genotype_matrix,         # (n_haploids, n_sites)
    positions=site_positions,    # (n_sites,) in bp
    model=model,
    blocks=[(0, 1_000_000), (1_000_000, 2_000_000)],
    pivot_pairs=[(0, 1)],
    mutation_rate=3.5e-9,
)

Universal entry point

The translate function auto-detects input type:

from fastcxt.translate import translate

# Accepts tree sequences or (gm, positions) tuples
means, variances, index_map = translate(ts, model, pivot_pairs=[(0, 1)])
means, variances, index_map = translate((gm, pos), model, pivot_pairs=[(0, 1)])

Understanding the output

• means: predicted log-TMRCA per window. Exponentiate for natural scale: np.exp(means)

• variances: predicted variance of log-TMRCA. Use for confidence intervals: np.exp(means ± 1.96 * np.sqrt(variances))

• index_map: maps each output row to [block_index, pair_index]

Scaling to whole genomes

For whole-genome analysis, use the TimeAtlas:

from fastcxt.atlas import TimeAtlas

atlas = TimeAtlas()
for arm in ["2L", "2R", "3L", "3R", "X"]:
    result = analysis.run_chromosome_arm(gm, pos, arm, pairs, mutation_rate)
    atlas.add_arm(arm, result["means"], result["variances"], pairs)
atlas.save("genome_atlas/")

See TimeAtlas for the full query API and Visualization for generating publication-quality figures from atlas data.