nmt.py 文件源码

def pred_probs(f_log_probs, prepare_data, options, iterator, verbose=True, normalize=False, alignweights=False):
    probs = []
    n_done = 0

    alignments_json = []

    for x, y in iterator:
        n_done += len(x)

        x, x_mask, y, y_mask = prepare_data(x, y,
                                            n_words_src=options['n_words_src'],
                                            n_words=options['n_words'])

        ### in optional save weights mode.
        if alignweights:
            pprobs, attention = f_log_probs(x, x_mask, y, y_mask)
            for jdata in get_alignments(attention, x_mask, y_mask):
                alignments_json.append(jdata)
        else:
            pprobs = f_log_probs(x, x_mask, y, y_mask)

        # normalize scores according to output length
        if normalize:
            lengths = numpy.array([numpy.count_nonzero(s) for s in y_mask.T])
            pprobs /= lengths

        for pp in pprobs:
            probs.append(pp)

        if numpy.isnan(numpy.mean(probs)):
            ipdb.set_trace()

        if verbose:
            print >>sys.stderr, '%d samples computed' % (n_done)

    return numpy.array(probs), alignments_json


# optimizers
# name(hyperp, tparams, grads, inputs (list), cost) = f_grad_shared, f_update