← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 6602.d · 生成模型:GPT-6

序列损失换算要数有效标签

理解等级:高级

若模型loss是有效标签上的平均负对数似然,应乘以同一批次中实际被计入的标签数量,才能恢复对应对数似然和。padding、忽略标签及decoder起始token都会改变计数;原始序列张量长度不一定等于有效预测数。

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

      log_probs = -loss * output_sequence.size(1) #2 else: raise ValueError("Unsupported model type") return torch.tensor(log_probs.item()) #1 Convert logits to logprobs for GPT-2. #2 Convert logits to logprobs from BART cross-entropy. Finally, we’ll put these pieces together to generate a full story from our three models. Listing 13.17 Generating a full story king_output = generate_from_model(king_model) #1 king_statement = convert_to_text(king_output) #1 print("Generated from king_nodel:", king_statement) #1 log_prob_king = compute_log_probs(king_model,

    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:摘录中关于序列损失换算要数有效标签的定义、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:log_probs = -loss * output_sequence.size(1) #2 else: raise ValueError("Unsupported model type") return torch.tensor(log_probs.item()) #1 Convert logits to logprobs for GPT-2. #2 Convert logits to logprobs from BART cross-entropy. Finally, we’ll put these pieces together to generate a full story from our three models. Listing 13.17 Generating a full story king_output = generate_from_model(king_model) #1 king_statement = convert_to_text(king_output) #1 print("Generated from king_nodel:", king_statement) #1 log_prob_king = compute_log_probs(king_model,