if beam_search: #beam search
            k = 10 #store best k options
            ## CONVERT BEST_OPTIONS WITH USE_GPU = TRUE OR FALSE.
            best_options = [(Variable(torch.zeros(1)).cuda(), Variable(torch.LongTensor([EN.vocab.stoi["<s>"]])).cuda(), states)]
            length = 0
            while length < max_trg_len:
                options = [] #same format as best_options
                for lprob, sentence, current_state in best_options:
                    last_word = sentence[-1]
                    if last_word.data[0] != EN.vocab.stoi["</s>"]:
                        probs, new_state = decoder(last_word.unsqueeze(1), current_state, outputs)
                        probs = probs.squeeze()
                        for index in torch.topk(probs, k)[1]: #only care about top k options in probs for next word.
                            options.append((torch.add(probs[index], lprob), torch.cat([sentence, index]), new_state))
                    else:
                        options.append((lprob, sentence, current_state))
                options.sort(key = lambda x: x[0].data[0], reverse=True)
                best_options = options[:k] #sorts by first element, which is lprob.
                length = length + 1
            best_options.sort(key = lambda x: x[0].data[0], reverse=True)
            best_choice = best_options[0] #best overall
            sentence = best_choice[1].data