知识卡片

普通读书笔记卡 · 6603.a · 生成模型:GPT-6

重采样索引必须对齐样本池

理解等级:中级

若按original_samples的权重抽索引,就必须用索引访问同一个有序样本池。先set去重并重排、仍使用原权重索引,会错配甚至越界。确需合并重复项时,要共同聚合其概率质量并重建对应权重。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      unique_samples = list(set(original_samples)) #2 log_importance_weights = torch.tensor( #2 [sample[2] for sample in original_samples]) #2 resampling_dist = Categorical(logits=log_importance_weights) #2 resampled_indices = resampling_dist.sample_n(num_samples) #2 samples = pd.DataFrame( #2 [unique_samples[i] for i in resampled_indices], #2 columns=["King", "Kingdom", "log_importance_weight"] #2 ) #2 samples["Prince"] = PRINCE_STORY samples["Distribution"] = "observational" return samples[['King', 'Prince', 'Kingdom', 'Distribution']] num_samples = 1000 posterior_samples = do_importance_resampling( cond_model, proposal_given_prince, num_samples) #1 Use importance resampling as our inference

    参考来源

    - 位置:《Causal AI》第13章《因果与大语言模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Causal AI/_epub-src/OEBPS/Text/chapter-13.xhtml) - 结论依据:摘录中关于重采样索引必须对齐样本池的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:unique_samples = list(set(original_samples)) #2 log_importance_weights = torch.tensor( #2 [sample[2] for sample in original_samples]) #2 resampling_dist = Categorical(logits=log_importance_weights) #2 resampled_indices = resampling_dist.sample_n(num_samples) #2 samples = pd.DataFrame( #2 [unique_samples[i] for i in resampled_indices], #2 columns=["King", "Kingdom", "log_importance_weight"] #2 ) #2 samples["Prince"] = PRINCE_STORY samples["Distribution"] = "observational" return samples[['King', 'Prince', 'Kingdom', 'Distribution']] num_samples = 1000 posterior_samples = do_importance_resampling( cond_model, proposal_given_prince, num_samples) #1 Use importance resampling as our inference