Pytorch 自定义激活函数前向与反向传播 Tanh

看完这篇，你基本上可以自定义前向与反向传播，可以自己定义自己的算子

import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

%matplotlib inline

plt.rcParams['figure.figsize'] = (7, 3.5)
plt.rcParams['figure.dpi'] = 150
plt.rcParams['axes.unicode_minus'] = False  #解决坐标轴负数的铅显示问题

Tanh

公式

$$\tanh (x)=\frac{\sinh (x)}{\cosh (x)}=\frac{e^x-e^{-x}}{e^x+e^{-x}}$$

$$\tanh (x)=2\sigma (2x)-1$$

求导过程

$\begin{array}{rl}{\tanh }^{\prime }(x)=& (\frac{e^x-e^{-x}}{e^x+e^{-x}}{)}^{\prime }\\ =& [(e^x-e^{-x})(e^x+e^{-x}{)}^{-1}{]}^{\prime }\\ =& (e^x+e^{-x})(e^x+e^{-x}{)}^{-1}+(e^x-e^{-x})(-1)(e^x+e^{-x}{)}^{-2}(e^x-e^{-x})\\ =& 1-(e^x-e^{-x}{)}^2(e^x+e^{-x}{)}^{-2}\\ =& 1-\frac{(e^x-e^{-x}{)}^2}{(e^x+e^{-x}{)}^2}\\ =& 1-{\tanh }^2(x)\end{array}$

优点：

Tanh也称为双切正切函数，取值范围为[-1,1]。tanh在特征相差明显时的效果会很好，在循环过程中会不断扩大特征效果。与 sigmoid 的区别是，tanh 是 0 均值的，因此实际应用中 tanh 会比 sigmoid 更好。文献 [LeCun, Y., et al., Backpropagation applied to handwritten zip code recognition. Neural computation, 1989. 1(4): p. 541-551.] 中提到tanh 网络的收敛速度要比sigmoid快，因为tanh 的输出均值比 sigmoid 更接近 0，SGD会更接近 natural gradient[4]（一种二次优化技术），从而降低所需的迭代次数。非常优秀，几乎适合所有的场景

缺点：

• 该导数在正负饱和区的梯度都会接近于0值，会造成梯度消失。还有其更复杂的幂运算。

自定义Tanh

class SelfDefinedTanh(torch.autograd.Function):
    @staticmethod
    def forward(ctx, inp):
        exp_x = torch.exp(inp)
        exp_x_ = torch.exp(-inp)
        result = torch.divide((exp_x - exp_x_), (exp_x + exp_x_))
        ctx.save_for_backward(result)
        return result

    @staticmethod
    def backward(ctx, grad_output):
        # ctx.saved_tensors is tuple (tensors, grad_fn)
        result, = ctx.saved_tensors
        return grad_output * (1 - result.pow(2))


class Tanh(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        out = SelfDefinedTanh.apply(x)
        return out

def tanh_sigmoid(x):
    """according to the equation"""
    # 2 * torch.sigmoid(2 * x) -1 
    return torch.mul(torch.sigmoid(torch.mul(x, 2)), 2) - 1

与Torch定义的比较

# self defined
torch.manual_seed(0)

tanh = Tanh()  # SelfDefinedTanh
inp = torch.randn(5, requires_grad=True)
out = tanh((inp + 1).pow(2))

print(f'Out is\n{out}')

out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nFirst call\n{inp.grad}")

out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nSecond call\n{inp.grad}")

inp.grad.zero_()
out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nCall after zeroing gradients\n{inp.grad}")

Out is
tensor([1.0000, 0.4615, 0.8831, 0.9855, 0.0071],
       grad_fn=<SelfDefinedTanhBackward>)

First call
tensor([ 5.0889e-05,  1.1121e+00, -5.1911e-01,  9.0267e-02, -1.6904e-01])

Second call
tensor([ 1.0178e-04,  2.2243e+00, -1.0382e+00,  1.8053e-01, -3.3807e-01])

Call after zeroing gradients
tensor([ 5.0889e-05,  1.1121e+00, -5.1911e-01,  9.0267e-02, -1.6904e-01])

# self defined tanh_sigmoid
torch.manual_seed(0)

inp = torch.randn(5, requires_grad=True)
out = tanh_sigmoid((inp + 1).pow(2))

print(f'Out is\n{out}')

out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nFirst call\n{inp.grad}")

out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nSecond call\n{inp.grad}")

inp.grad.zero_()
out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nCall after zeroing gradients\n{inp.grad}")

Out is
tensor([1.0000, 0.4615, 0.8831, 0.9855, 0.0071], grad_fn=<SubBackward0>)

First call
tensor([ 5.0889e-05,  1.1121e+00, -5.1911e-01,  9.0267e-02, -1.6904e-01])

Second call
tensor([ 1.0178e-04,  2.2243e+00, -1.0382e+00,  1.8053e-01, -3.3807e-01])

Call after zeroing gradients
tensor([ 5.0889e-05,  1.1121e+00, -5.1911e-01,  9.0267e-02, -1.6904e-01])

# torch defined
torch.manual_seed(0)

inp = torch.randn(5, requires_grad=True)
out = torch.tanh((inp + 1).pow(2))

print(f'Out is\n{out}')

out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nFirst call\n{inp.grad}")

out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nSecond call\n{inp.grad}")

inp.grad.zero_()
out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nCall after zeroing gradients\n{inp.grad}")

Out is
tensor([1.0000, 0.4615, 0.8831, 0.9855, 0.0071], grad_fn=<TanhBackward>)

First call
tensor([ 5.0283e-05,  1.1121e+00, -5.1911e-01,  9.0267e-02, -1.6904e-01])

Second call
tensor([ 1.0057e-04,  2.2243e+00, -1.0382e+00,  1.8053e-01, -3.3807e-01])

Call after zeroing gradients
tensor([ 5.0283e-05,  1.1121e+00, -5.1911e-01,  9.0267e-02, -1.6904e-01])

从上3个结果，可以看出，不管是经过sigmoid来计算，还是公式定义都可以得到一样的output与gradient。但在输入的值较大时，torch应该是减去一个小值，使得梯度更小。

可视化

# visualization
inp = torch.arange(-8, 8, 0.1, requires_grad=True)
out = tanh(inp)
out.sum().backward()

inp_grad = inp.grad

plt.plot(inp.detach().numpy(),
         out.detach().numpy(),
         label=r"$\tanh(x)$",
         alpha=0.7)
plt.plot(inp.detach().numpy(),
         inp_grad.numpy(),
         label=r"$\tanh'(x)$",
         alpha=0.5)
plt.grid()
plt.legend()
plt.show()