知识卡片
反向传播与参数更新要分开
loss.backward按计算图链式求导,将梯度积累到参数;optimizer.step按优化规则更新参数。两者职责不同,zero_grad用于在不需累积时清理旧梯度;不能把step本身叫反向传播,也不能省略梯度清理而误以为仍是每轮独立更新。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 神经网络基础
原书出处 · 1 处
- · 查看原书依据
... optimizer.zero_grad() #1 ... outputs = model(X_train) ... loss_train = loss_func_train(outputs, y_train) #2 ... loss_train.backward() #3 ... optimizer.step() #4 Epoch:: 100%|================================| 200/200 [00:04<00:00, 42.84it/s] ↪ 96.26it/s] #1 Step 1: Sets the stored gradients to 0 #2 Step 2: Calculates the training loss #3 Step 3: Computes the gradient on the training set #4 Step 4: Updates weights and biases with the optimizer (backprop) That was
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第05章《神经网络基础》(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-5.html)
- 结论依据:摘录中关于反向传播与参数更新要分开的定义、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:... optimizer.zero_grad() #1 ... outputs = model(X_train) ... loss_train = loss_func_train(outputs, y_train) #2 ... loss_train.backward() #3 ... optimizer.step() #4 Epoch:: 100%|================================| 200/200 [00:04<00:00, 42.84it/s] ↪ 96.26it/s] #1 Step 1: Sets the stored gradients to 0 #2 Step 2: Calculates the training loss #3 Step 3: Computes the gradient on the training set #4 Step 4: Updates weights and biases with the optimizer (backprop) That was