刘二老师第7节,多个激活器图例
由于编码能力有限,略显简陋,大牛勿喷
Model.py. 主类文件
import torch
import numpy as np
import matplotlib.pyplot as plt
from activate import activate_fun
# data_set
xy = np.loadtxt('diabetes.csv.gz', delimiter=',', dtype=np.float32) # 第二个参数相当于分割作用
x_data = torch.from_numpy(xy[:, :-1]) # (所有行,从第一行开始,最后一行不要)
y_data = torch.from_numpy(xy[:, [-1]]) # 用[]保证拿出来的是个矩阵而不是向量 (所有行,只要最后一列)
activate_f = [
activate_fun('ReLU6'),
activate_fun('ReLU'),
activate_fun('ELU'),
activate_fun('PReLU'),
activate_fun('LeakyReLU'),
activate_fun('Threshold', 0.2, 0.0),
activate_fun('Hardtanh'),
activate_fun('Sigmoid'),
activate_fun('Tanh'),
activate_fun('LogSigmoid'),
activate_fun('Softplus'),
activate_fun('Softshrink'),
activate_fun('Softsign'),
activate_fun('Softmin'),
activate_fun('Softmax'),
activate_fun('LogSoftmax')
]
# model class
class multipleLogistic(torch.nn.Module):
def __init__(self, act):
super(multipleLogistic, self).__init__()
self.linear1 = torch.nn.Linear(8, 6)
self.linear2 = torch.nn.Linear(6, 4)
self.linear3 = torch.nn.Linear(4, 1) # 加了3个层
self.activate = act # 继承的仍然还是module模块
self.sigmoid = torch.nn.Sigmoid()
def forward(self, x):
x = self.activate(self.linear1(x)) # 第一层的y_pred经过激活函数输出
x = self.activate(self.linear2(x)) # 第二层输入为第一层输出
x = self.activate(self.linear3(x))
x = self.sigmoid(x)
return x
epoch_list = range(1, 101)
loss_list = [[] for _ in activate_f]
for act_f, i in zip(activate_f, range(len(activate_f))):
model = multipleLogistic(act_f)
# criterion optimizer
criterion = torch.nn.BCELoss(size_average=True)
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
# training model
for epoch in range(100):
y_pred = model(x_data)
loss = criterion(y_pred, y_data)
loss_list[i].append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
fig, ax = plt.subplots()
ax.plot(epoch_list, loss_list[0], label='ReLU6')
ax.plot(epoch_list, loss_list[1], label='ReLU')
ax.plot(epoch_list, loss_list[2], label='ELU')
ax.plot(epoch_list, loss_list[3], label='PReLU')
ax.plot(epoch_list, loss_list[4], label='LeakyReLU')
ax.plot(epoch_list, loss_list[5], label='Threshold')
ax.plot(epoch_list, loss_list[6], label='Hardtanh')
ax.plot(epoch_list, loss_list[7], label='Sigmoid')
ax.plot(epoch_list, loss_list[8], label='Tanh')
ax.plot(epoch_list, loss_list[9], label='LogSigmoid')
ax.plot(epoch_list, loss_list[10], label='Softplus')
ax.plot(epoch_list, loss_list[11], label='Softshrink')
ax.plot(epoch_list, loss_list[12], label='Softsign')
ax.plot(epoch_list, loss_list[13], label='Softmin')
ax.plot(epoch_list, loss_list[14], label='Softmax')
ax.plot(epoch_list, loss_list[15], label='LogSoftmax')
ax.set_xlabel('rounds')
ax.set_ylabel('loss')
ax.set_title('activate_fun')
ax.legend()
plt.savefig('p.png')
plt.show()
activity.py 激活函数文件
import torch
def activate_fun(fun_name, *args, **kwargs):
if fun_name == 'ReLU6':
return torch.nn.ReLU6(*args, **kwargs)
elif fun_name == 'ReLU':
return torch.nn.ReLU(*args, **kwargs)
elif fun_name == 'ELU':
return torch.nn.ELU(*args, **kwargs)
elif fun_name == 'PReLU':
return torch.nn.PReLU(*args, **kwargs)
elif fun_name == 'LeakyReLU':
return torch.nn.LeakyReLU(*args, **kwargs)
elif fun_name == 'Threshold':
return torch.nn.Threshold(*args, **kwargs)
elif fun_name == 'Hardtanh':
return torch.nn.Hardtanh(*args, **kwargs)
elif fun_name == 'Sigmoid':
return torch.nn.Sigmoid()
elif fun_name == 'Tanh':
return torch.nn.Tanh()
elif fun_name == 'LogSigmoid':
return torch.nn.LogSigmoid()
elif fun_name == 'Softplus':
return torch.nn.Softplus(*args, **kwargs)
elif fun_name == 'Softshrink':
return torch.nn.Softshrink(*args, **kwargs)
elif fun_name == 'Softsign':
return torch.nn.Softsign()
elif fun_name == 'Softmin':
return torch.nn.Softmin()
elif fun_name == 'Softmax':
return torch.nn.Softmax(dim=1)
elif fun_name == 'LogSoftmax':
return torch.nn.LogSoftmax(dim=1)
else:
raise ValueError(f"不支持的激活函数: {fun_name}")
由于激活函数许多有的概率映射无法到0-1范围,所以程序运算时按照刘老师的源码会出问题,所以我在模型类里初始函数和前馈函数中神经网络最后一层加了sigmoid函数,将之前的激活函数输出值在映射一次到0-1范围内,不知道这种方法是否可行,不过我的目的是为了看各种激活函数的效果,所以具体影响就没有细究。以下是图例:
P S:有很多不会的,必要时确实需要问gpt,但不能盲从