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

def info_nce(q, k, tau=0.05):
    # q, k: (N, d). Row i of q matches row i of k.
    q, k = np.asarray(q, dtype=float), np.asarray(k, dtype=float)
    if q.ndim != 2 or q.shape != k.shape or len(q) == 0 or tau <= 0:
        raise ValueError("Use matching nonempty matrices and positive temperature")
    if (np.linalg.norm(q, axis=1) == 0).any() or (np.linalg.norm(k, axis=1) == 0).any():
        raise ValueError("Cosine similarity needs nonzero vectors")
    q = q / np.linalg.norm(q, axis=1, keepdims=True)
    k = k / np.linalg.norm(k, axis=1, keepdims=True)
    logits = q @ k.T / tau                          # (N, N)
    logits -= logits.max(axis=1, keepdims=True)
    log_prob = logits - np.log(np.exp(logits).sum(axis=1, keepdims=True))
    return -np.mean(np.diag(log_prob))              # positives on the diagonal
