知识卡片

普通读书笔记卡 · 6733 · 生成模型:GPT-6

末步损失会经过全部相关时间步

理解等级:中级

对完整序列前向后,以末步输出算损失并一次backward,梯度可通过隐藏依赖回到更早位置。它不要求每个词都各自调用参数更新;若只在最后监督,需确认序列非空且训练图未被意外切断。

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

      >>> def train_sample(model, category_tensor, char_seq_tens, ... criterion=nn.NLLLoss(), lr=.005): """ Train for one epoch (one example name nationality tensor pair) """ ... hidden = torch.zeros(1, model.n_hidden) #1 ... model.zero_grad() #2 ... for char_onehot_vector in char_seq_tens: ... category_predictions, hidden = model( #3 ... x=char_onehot_vector, hidden=hidden) #4 ... loss = criterion(category_predictions, category_tensor) ... loss.backward() ... for p in model.parameters(): ... p.data.add_(p.grad.data, alpha=-lr) ... return model, category_predictions, loss.item() #1

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-8.html) - 结论依据:摘录中关于末步损失会经过全部相关时间步的定义、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:>>> def train_sample(model, category_tensor, char_seq_tens, ... criterion=nn.NLLLoss(), lr=.005): """ Train for one epoch (one example name nationality tensor pair) """ ... hidden = torch.zeros(1, model.n_hidden) #1 ... model.zero_grad() #2 ... for char_onehot_vector in char_seq_tens: ... category_predictions, hidden = model( #3 ... x=char_onehot_vector, hidden=hidden) #4 ... loss = criterion(category_predictions, category_tensor) ... loss.backward() ... for p in model.parameters(): ... p.data.add_(p.grad.data, alpha=-lr) ... return model, category_predictions, loss.item() #1