#!/usr/bin/env python3
# Source: q3-deep-learning.html
import numpy as np

def sample_top_p(logits, p=0.9, temperature=0.7, rng=np.random):
    logits = np.asarray(logits, dtype=float)
    if logits.ndim != 1 or logits.size == 0 or not np.isfinite(logits).all():
        raise ValueError("Use nonempty finite one-dimensional logits")
    if not 0 < p <= 1 or temperature < 0:
        raise ValueError("Use 0 < p <= 1 and nonnegative temperature")
    if temperature == 0:
        return int(np.argmax(logits))
    z = logits / temperature
    probs = np.exp(z - z.max())
    probs /= probs.sum()
    order = np.argsort(-probs)                 # most likely first
    cum = np.cumsum(probs[order])
    keep = order[: np.searchsorted(cum, p) + 1]  # smallest set reaching p
    q = probs[keep] / probs[keep].sum()
    return rng.choice(keep, p=q)
