526互联

nlp基础-生成模型解码策略

发布时间 2023-10-13 14:58:34作者: shiiiilong

首先参考transformers的源代码

# transformers.generation.utils..GenerationMixin._get_generation_mode
if generation_config.constraints is not None or generation_config.force_words_ids is not None:
            generation_mode = GenerationMode.CONSTRAINED_BEAM_SEARCH
        elif generation_config.num_beams == 1:
            if generation_config.do_sample is False:
                if (
                    generation_config.top_k is not None
                    and generation_config.top_k > 1
                    and generation_config.penalty_alpha is not None
                    and generation_config.penalty_alpha > 0
                ):
                    generation_mode = GenerationMode.CONTRASTIVE_SEARCH
                else:
                    generation_mode = GenerationMode.GREEDY_SEARCH
            else:
                generation_mode = GenerationMode.SAMPLE
        else:
            if generation_config.num_beam_groups > 1:
                generation_mode = GenerationMode.GROUP_BEAM_SEARCH
            elif generation_config.do_sample is True:
                generation_mode = GenerationMode.BEAM_SAMPLE
            else:
                generation_mode = GenerationMode.BEAM_SEARCH