def evaluate()

in lib/metrics.py [0:0]

• 33 lines of code
• 7 McCabe index (conditional complexity)

def evaluate(net, xq, xb, gt, quantizers, best_key, device=None,
             trainset=None):
    net.eval()
    if device is None:
        device = next(net.parameters()).device.type
    xqt = forward_pass(net, sanitize(xq), device=device)
    xbt = forward_pass(net, sanitize(xb), device=device)
    if trainset is not None:
        trainset = forward_pass(net, sanitize(trainset), device=device)
    nq, d = xqt.shape
    res = {}
    score = 0
    for quantizer in quantizers:
        qt = getQuantizer(quantizer, d)

        qt.train(trainset)
        xbtq = qt(xbt)
        if not qt.asymmetric:
            xqtq = qt(xqt)
            I = get_nearestneighbors(xqtq, xbtq, 100, device)
        else:
            I = get_nearestneighbors(xqt, xbtq, 100, device)

        print("%s\t nbit=%3d: " % (quantizer, qt.bits), end=' ')

        # compute 1-recall at ranks 1, 10, 100 (comparable with
        # fig 5, left of the paper)
        recalls = []
        for rank in 1, 10, 100:
            recall = (I[:, :rank] == gt[:, :1]).sum() / float(nq)
            key = '%s,rank=%d' % (quantizer, rank)
            if key == best_key:
                score = recall
            recalls.append(recall)
            print('%.4f' % recall, end=" ")
        res[quantizer] = recalls
        print("")

    return res, score