其他分享
首页 > 其他分享> > 深度学习的前向传播/反向传播总结

深度学习的前向传播/反向传播总结

作者:互联网

以下内容根据cs231n课程材料总结。

文章目录

0. 参考资料

https://cs231n.github.io/optimization-2/
http://cs231n.stanford.edu/handouts/derivatives.pdf
http://cs231n.stanford.edu/slides/2021/lecture_4.pdf
cs231n 作业2

1. 全连接层

在这里插入图片描述

前向传播

Y = X W + B (1) Y=XW+B\tag{1} Y=XW+B(1)

def affine_forward(x, w, b):
    out = None # 这一层前向传播的结果,也就是上面的Y
    
    N = x.shape[0]
    x_ = x.reshape(N, -1) # shape: (N, D1, D2, ···,Dk) => (N, D)
    out = x_ @ w + b
    
    cache = (x, w, b) # 用于反向传播
    return out, cache

反向传播


Y n , m = Z n , m + B m (2) Y_{n,m}=Z_{n,m}+B_{m}\tag{2} Yn,m​=Zn,m​+Bm​(2)
∂ L ∂ B m = ∑ n ∂ L ∂ Y n , m ∂ Y n , m ∂ B m = ∑ n ∂ L ∂ Y n , m ⋅ 1 = ∑ n ∂ L ∂ Y n , m (3) \begin{aligned}\frac{\partial{L}}{\partial{B_{m}}}&=\sum_{n}{\frac{\partial{L}}{\partial{Y_{n,m}}}\frac{\partial{Y_{n,m}}}{\partial{B_{m}}}}\\&=\sum_{n}{\frac{\partial{L}}{\partial{Y_{n,m}}}\cdot1}\\&=\sum_{n}{\frac{\partial{L}}{\partial{Y_{n,m}}}}\end{aligned}\tag{3} ∂Bm​∂L​​=n∑​∂Yn,m​∂L​∂Bm​∂Yn,m​​=n∑​∂Yn,m​∂L​⋅1=n∑​∂Yn,m​∂L​​(3)

∂ L ∂ B = ∑ n ∂ L ∂ Y n , ⋅ (4) \frac{\partial{L}}{\partial{B}}=\sum_n{\frac{\partial{L}}{\partial{Y_{n,\cdot}}}}\tag{4} ∂B∂L​=n∑​∂Yn,⋅​∂L​(4)
∂ L ∂ Z n , m = ∂ L ∂ Y n , m ∂ Y n , m ∂ Z n , m = ∂ L ∂ Y n , m ⋅ 1 = ∂ L ∂ Y n , m (5) \begin{aligned}\frac{\partial{L}}{\partial{Z_{n,m}}}&=\frac{\partial{L}}{\partial{Y_{n,m}}}\frac{\partial{Y_{n,m}}}{\partial{Z_{n,m}}}\\&=\frac{\partial{L}}{\partial{Y_{n,m}}}\cdot1\\&=\frac{\partial{L}}{\partial{Y_{n,m}}}\end{aligned}\tag{5} ∂Zn,m​∂L​​=∂Yn,m​∂L​∂Zn,m​∂Yn,m​​=∂Yn,m​∂L​⋅1=∂Yn,m​∂L​​(5)

∂ L ∂ Z = ∂ L ∂ Y (6) \frac{\partial{L}}{\partial{Z}}=\frac{\partial{L}}{\partial{Y}}\tag{6} ∂Z∂L​=∂Y∂L​(6)
由矩阵乘法公式可知
Z n , m = ∑ d X n , d W d , m (7) Z_{n,m}=\sum_d{X_{n,d}W_{d,m}}\tag{7} Zn,m​=d∑​Xn,d​Wd,m​(7)
∂ L ∂ X n , d = ∑ m ∂ L ∂ Z n , m ∂ Z n , m ∂ X n , d = ∑ m ∂ L ∂ Z n , m W d , m (8) \begin{aligned}\frac{\partial{L}}{\partial{X_{n,d}}}&=\sum_m{\frac{\partial{L}}{\partial{Z_{n,m}}}\frac{\partial{Z_{n,m}}}{\partial{X_{n,d}}}}\\&=\sum_m{\frac{\partial{L}}{\partial{Z_{n,m}}}W_{d,m}}\end{aligned}\tag{8} ∂Xn,d​∂L​​=m∑​∂Zn,m​∂L​∂Xn,d​∂Zn,m​​=m∑​∂Zn,m​∂L​Wd,m​​(8)

∂ L ∂ X = ∂ L ∂ Z W T (9) \frac{\partial{L}}{\partial{X}}=\frac{\partial{L}}{\partial{Z}}W^T\tag{9} ∂X∂L​=∂Z∂L​WT(9)
∂ L ∂ W d , m = ∑ n ∂ L ∂ Z n , m ∂ Z n , m ∂ W d , m = ∑ n ∂ L ∂ Z n , m X n , d (10) \begin{aligned}\frac{\partial{L}}{\partial{W_{d,m}}}&=\sum_n{\frac{\partial{L}}{\partial{Z_{n,m}}}\frac{\partial{Z_{n,m}}}{\partial{W_{d,m}}}}\\&=\sum_n{\frac{\partial{L}}{\partial{Z_{n,m}}}X_{n,d}}\end{aligned}\tag{10} ∂Wd,m​∂L​​=n∑​∂Zn,m​∂L​∂Wd,m​∂Zn,m​​=n∑​∂Zn,m​∂L​Xn,d​​(10)

∂ L ∂ W = X T ∂ L ∂ Z (11) \frac{\partial{L}}{\partial{W}}=X^T\frac{\partial{L}}{\partial{Z}}\tag{11} ∂W∂L​=XT∂Z∂L​(11)

def affine_backward(dout, cache):
    """Computes the backward pass for an affine (fully connected) layer.

    Inputs:
    - dout: Upstream derivative, of shape (N, M)
    - cache: Tuple of:
      - x: Input data, of shape (N, d_1, ... d_k)
      - w: Weights, of shape (D, M)
      - b: Biases, of shape (M,)

    Returns a tuple of:
    - dx: Gradient with respect to x, of shape (N, d1, ..., d_k)
    - dw: Gradient with respect to w, of shape (D, M)
    - db: Gradient with respect to b, of shape (M,)
    """
    x, w, b = cache
    dx, dw, db = None, None, None
    
    N = x.shape[0]
    x_ = x.reshape(N, -1)
    db = np.sum(dout, axis=0)
    dw = x_.T @ dout 
    dx = dout @ w.T

    return dx, dw, db

2. ReLU

在这里插入图片描述

前向传播

Inputs: 
	- X: (N, D)

Returns:
	- Y: (N, D)

Y = max ⁡ ( 0 , X ) (12) Y=\max{(0,X)}\tag{12} Y=max(0,X)(12)

def relu_forward(x):
    out = None

    out = np.maximum(0, x)

    cache = x
    return out, cache

反向传播


∂ L ∂ X n , d = ∂ L ∂ Y n , d ∂ Y n , d ∂ X n , d = ∂ L ∂ Y n , d 1 { X n , d > 0 } (13) \begin{aligned}\frac{\partial{L}}{\partial{X_{n,d}}}&=\frac{\partial{L}}{\partial{Y_{n,d}}}\frac{\partial{Y_{n,d}}}{\partial{X_{n,d}}}\\&=\frac{\partial{L}}{\partial{Y_{n,d}}}\mathbf{1}\{X_{n,d}>0\}\end{aligned}\tag{13} ∂Xn,d​∂L​​=∂Yn,d​∂L​∂Xn,d​∂Yn,d​​=∂Yn,d​∂L​1{Xn,d​>0}​(13)

def relu_backward(dout, cache):
    """Computes the backward pass for a layer of rectified linear units (ReLUs).

    Input:
    - dout: Upstream derivatives, of any shape
    - cache: Input x, of same shape as dout

    Returns:
    - dx: Gradient with respect to x
    """
    dx, x = None, cache

    dx = (x > 0) * dout  
    
    return dx

标签:partial,Yn,cache,传播,shape,tag,反向,深度,frac
来源: https://blog.csdn.net/pgsld2333/article/details/122494870