佛山 建站公司,佛山企业网站设计公司,网站服务搭建,论客企业邮箱官网9.1 优化器 ① 损失函数调用backward方法#xff0c;就可以调用损失函数的反向传播方法#xff0c;就可以求出我们需要调节的梯度#xff0c;我们就可以利用我们的优化器就可以根据梯度对参数进行调整#xff0c;达到整体误差降低的目的。 ② 梯度要清零#xff0c;如果梯…9.1 优化器 ① 损失函数调用backward方法就可以调用损失函数的反向传播方法就可以求出我们需要调节的梯度我们就可以利用我们的优化器就可以根据梯度对参数进行调整达到整体误差降低的目的。 ② 梯度要清零如果梯度不清零会导致梯度累加。 9.2 神经网络优化一轮
import torch
import torchvision
from torch import nn
from torch.nn import Conv2d, MaxPool2d, Flatten, Linear, Sequential
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriterdataset torchvision.datasets.CIFAR10(./dataset,trainFalse,transformtorchvision.transforms.ToTensor(),downloadTrue)
dataloader DataLoader(dataset, batch_size64,drop_lastTrue)class Tudui(nn.Module):def __init__(self):super(Tudui, self).__init__() self.model1 Sequential(Conv2d(3,32,5,padding2),MaxPool2d(2),Conv2d(32,32,5,padding2),MaxPool2d(2),Conv2d(32,64,5,padding2),MaxPool2d(2),Flatten(),Linear(1024,64),Linear(64,10))def forward(self, x):x self.model1(x)return xloss nn.CrossEntropyLoss() # 交叉熵
tudui Tudui()
optim torch.optim.SGD(tudui.parameters(),lr0.01) # 随机梯度下降优化器
for data in dataloader:imgs, targets dataoutputs tudui(imgs)result_loss loss(outputs, targets) # 计算实际输出与目标输出的差距optim.zero_grad() # 梯度清零result_loss.backward() # 反向传播计算损失函数的梯度optim.step() # 根据梯度对网络的参数进行调优print(result_loss) # 对数据只看了一遍只看了一轮所以loss下降不大
结果 Files already downloaded and verified
tensor(2.2978, grad_fnNllLossBackward0)
tensor(2.2988, grad_fnNllLossBackward0)
tensor(2.3163, grad_fnNllLossBackward0)
tensor(2.3253, grad_fnNllLossBackward0)
tensor(2.2952, grad_fnNllLossBackward0)
tensor(2.3066, grad_fnNllLossBackward0)
tensor(2.3085, grad_fnNllLossBackward0)
tensor(2.3106, grad_fnNllLossBackward0)
tensor(2.2960, grad_fnNllLossBackward0)
tensor(2.3053, grad_fnNllLossBackward0)
tensor(2.2892, grad_fnNllLossBackward0)
tensor(2.3090, grad_fnNllLossBackward0)
tensor(2.2956, grad_fnNllLossBackward0)
tensor(2.3041, grad_fnNllLossBackward0)
tensor(2.3012, grad_fnNllLossBackward0)
tensor(2.3043, grad_fnNllLossBackward0)
tensor(2.2760, grad_fnNllLossBackward0)
tensor(2.3051, grad_fnNllLossBackward0)
tensor(2.2951, grad_fnNllLossBackward0)
tensor(2.3168, grad_fnNllLossBackward0)
tensor(2.3140, grad_fnNllLossBackward0)
tensor(2.3096, grad_fnNllLossBackward0)
tensor(2.2945, grad_fnNllLossBackward0)
tensor(2.3115, grad_fnNllLossBackward0)
tensor(2.2987, grad_fnNllLossBackward0)
tensor(2.3029, grad_fnNllLossBackward0)
tensor(2.3096, grad_fnNllLossBackward0)
tensor(2.3064, grad_fnNllLossBackward0)
tensor(2.3161, grad_fnNllLossBackward0)
tensor(2.3129, grad_fnNllLossBackward0)
tensor(2.2903, grad_fnNllLossBackward0)
tensor(2.3043, grad_fnNllLossBackward0)
tensor(2.3034, grad_fnNllLossBackward0)
tensor(2.3169, grad_fnNllLossBackward0)
tensor(2.3090, grad_fnNllLossBackward0)
tensor(2.3039, grad_fnNllLossBackward0)
tensor(2.3019, grad_fnNllLossBackward0)
tensor(2.3071, grad_fnNllLossBackward0)
tensor(2.3018, grad_fnNllLossBackward0)
tensor(2.3083, grad_fnNllLossBackward0)
tensor(2.2994, grad_fnNllLossBackward0)
tensor(2.2909, grad_fnNllLossBackward0)
tensor(2.3130, grad_fnNllLossBackward0)
tensor(2.2993, grad_fnNllLossBackward0)
tensor(2.2906, grad_fnNllLossBackward0)
tensor(2.3084, grad_fnNllLossBackward0)
tensor(2.3123, grad_fnNllLossBackward0)
tensor(2.2931, grad_fnNllLossBackward0)
tensor(2.3059, grad_fnNllLossBackward0)
tensor(2.3117, grad_fnNllLossBackward0)
tensor(2.2975, grad_fnNllLossBackward0)
tensor(2.3109, grad_fnNllLossBackward0)
tensor(2.3029, grad_fnNllLossBackward0)
tensor(2.3020, grad_fnNllLossBackward0)
tensor(2.3022, grad_fnNllLossBackward0)
tensor(2.3005, grad_fnNllLossBackward0)
tensor(2.2920, grad_fnNllLossBackward0)
tensor(2.3016, grad_fnNllLossBackward0)
tensor(2.3053, grad_fnNllLossBackward0)
tensor(2.3082, grad_fnNllLossBackward0)
tensor(2.3011, grad_fnNllLossBackward0)
tensor(2.3040, grad_fnNllLossBackward0)
tensor(2.3130, grad_fnNllLossBackward0)
tensor(2.2981, grad_fnNllLossBackward0)
tensor(2.2977, grad_fnNllLossBackward0)
tensor(2.2994, grad_fnNllLossBackward0)
tensor(2.3075, grad_fnNllLossBackward0)
tensor(2.3016, grad_fnNllLossBackward0)
tensor(2.2966, grad_fnNllLossBackward0)
tensor(2.3015, grad_fnNllLossBackward0)
tensor(2.3000, grad_fnNllLossBackward0)
tensor(2.2953, grad_fnNllLossBackward0)
tensor(2.2958, grad_fnNllLossBackward0)
tensor(2.2977, grad_fnNllLossBackward0)
tensor(2.2928, grad_fnNllLossBackward0)
tensor(2.2989, grad_fnNllLossBackward0)
tensor(2.2968, grad_fnNllLossBackward0)
tensor(2.2982, grad_fnNllLossBackward0)
tensor(2.2912, grad_fnNllLossBackward0)
tensor(2.3005, grad_fnNllLossBackward0)
tensor(2.2909, grad_fnNllLossBackward0)
tensor(2.2940, grad_fnNllLossBackward0)
tensor(2.2959, grad_fnNllLossBackward0)
tensor(2.2993, grad_fnNllLossBackward0)
tensor(2.2933, grad_fnNllLossBackward0)
tensor(2.2951, grad_fnNllLossBackward0)
tensor(2.2824, grad_fnNllLossBackward0)
tensor(2.2987, grad_fnNllLossBackward0)
tensor(2.2961, grad_fnNllLossBackward0)
tensor(2.2914, grad_fnNllLossBackward0)
tensor(2.3025, grad_fnNllLossBackward0)
tensor(2.2895, grad_fnNllLossBackward0)
tensor(2.2943, grad_fnNllLossBackward0)
tensor(2.2974, grad_fnNllLossBackward0)
tensor(2.2977, grad_fnNllLossBackward0)
tensor(2.3069, grad_fnNllLossBackward0)
tensor(2.2972, grad_fnNllLossBackward0)
tensor(2.2979, grad_fnNllLossBackward0)
tensor(2.2932, grad_fnNllLossBackward0)
tensor(2.2940, grad_fnNllLossBackward0)
tensor(2.3014, grad_fnNllLossBackward0)
tensor(2.2958, grad_fnNllLossBackward0)
tensor(2.3013, grad_fnNllLossBackward0)
tensor(2.2953, grad_fnNllLossBackward0)
tensor(2.2951, grad_fnNllLossBackward0)
tensor(2.3116, grad_fnNllLossBackward0)
tensor(2.2916, grad_fnNllLossBackward0)
tensor(2.2871, grad_fnNllLossBackward0)
tensor(2.2975, grad_fnNllLossBackward0)
tensor(2.2950, grad_fnNllLossBackward0)
tensor(2.3039, grad_fnNllLossBackward0)
tensor(2.2901, grad_fnNllLossBackward0)
tensor(2.2950, grad_fnNllLossBackward0)
tensor(2.2958, grad_fnNllLossBackward0)
tensor(2.2893, grad_fnNllLossBackward0)
tensor(2.2917, grad_fnNllLossBackward0)
tensor(2.3001, grad_fnNllLossBackward0)
tensor(2.2988, grad_fnNllLossBackward0)
tensor(2.3069, grad_fnNllLossBackward0)
tensor(2.3083, grad_fnNllLossBackward0)
tensor(2.2841, grad_fnNllLossBackward0)
tensor(2.2932, grad_fnNllLossBackward0)
tensor(2.2857, grad_fnNllLossBackward0)
tensor(2.2971, grad_fnNllLossBackward0)
tensor(2.2999, grad_fnNllLossBackward0)
tensor(2.2911, grad_fnNllLossBackward0)
tensor(2.2977, grad_fnNllLossBackward0)
tensor(2.3027, grad_fnNllLossBackward0)
tensor(2.2940, grad_fnNllLossBackward0)
tensor(2.2939, grad_fnNllLossBackward0)
tensor(2.2950, grad_fnNllLossBackward0)
tensor(2.2951, grad_fnNllLossBackward0)
tensor(2.3000, grad_fnNllLossBackward0)
tensor(2.2935, grad_fnNllLossBackward0)
tensor(2.2817, grad_fnNllLossBackward0)
tensor(2.2977, grad_fnNllLossBackward0)
tensor(2.3067, grad_fnNllLossBackward0)
tensor(2.2742, grad_fnNllLossBackward0)
tensor(2.2964, grad_fnNllLossBackward0)
tensor(2.2927, grad_fnNllLossBackward0)
tensor(2.2941, grad_fnNllLossBackward0)
tensor(2.3003, grad_fnNllLossBackward0)
tensor(2.2965, grad_fnNllLossBackward0)
tensor(2.2908, grad_fnNllLossBackward0)
tensor(2.2885, grad_fnNllLossBackward0)
tensor(2.2984, grad_fnNllLossBackward0)
tensor(2.3009, grad_fnNllLossBackward0)
tensor(2.2931, grad_fnNllLossBackward0)
tensor(2.2856, grad_fnNllLossBackward0)
tensor(2.2907, grad_fnNllLossBackward0)
tensor(2.2938, grad_fnNllLossBackward0)
tensor(2.2880, grad_fnNllLossBackward0)
tensor(2.2975, grad_fnNllLossBackward0)
tensor(2.2922, grad_fnNllLossBackward0)
tensor(2.2966, grad_fnNllLossBackward0)
tensor(2.2804, grad_fnNllLossBackward0) 9.3 神经网络优化多轮
import torch
import torchvision
from torch import nn
from torch.nn import Conv2d, MaxPool2d, Flatten, Linear, Sequential
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriterdataset torchvision.datasets.CIFAR10(./dataset,trainFalse,transformtorchvision.transforms.ToTensor(),downloadTrue)
dataloader DataLoader(dataset, batch_size64,drop_lastTrue)class Tudui(nn.Module):def __init__(self):super(Tudui, self).__init__() self.model1 Sequential(Conv2d(3,32,5,padding2),MaxPool2d(2),Conv2d(32,32,5,padding2),MaxPool2d(2),Conv2d(32,64,5,padding2),MaxPool2d(2),Flatten(),Linear(1024,64),Linear(64,10))def forward(self, x):x self.model1(x)return xloss nn.CrossEntropyLoss() # 交叉熵
tudui Tudui()
optim torch.optim.SGD(tudui.parameters(),lr0.01) # 随机梯度下降优化器
for epoch in range(20):running_loss 0.0for data in dataloader:imgs, targets dataoutputs tudui(imgs)result_loss loss(outputs, targets) # 计算实际输出与目标输出的差距optim.zero_grad() # 梯度清零result_loss.backward() # 反向传播计算损失函数的梯度optim.step() # 根据梯度对网络的参数进行调优running_loss running_loss result_lossprint(running_loss) # 对这一轮所有误差的总和
结果 Files already downloaded and verified
tensor(358.1069, grad_fnAddBackward0)
tensor(353.8411, grad_fnAddBackward0)
tensor(337.3790, grad_fnAddBackward0)
tensor(317.3237, grad_fnAddBackward0)
tensor(307.6762, grad_fnAddBackward0)
tensor(298.2425, grad_fnAddBackward0)
tensor(289.7010, grad_fnAddBackward0)
tensor(282.7116, grad_fnAddBackward0)
tensor(275.8972, grad_fnAddBackward0)
tensor(269.5961, grad_fnAddBackward0)
tensor(263.8480, grad_fnAddBackward0)
tensor(258.5006, grad_fnAddBackward0)
tensor(253.4671, grad_fnAddBackward0)
tensor(248.7994, grad_fnAddBackward0)
tensor(244.4917, grad_fnAddBackward0)
tensor(240.5728, grad_fnAddBackward0)
tensor(236.9719, grad_fnAddBackward0)
tensor(233.6264, grad_fnAddBackward0)
tensor(230.4298, grad_fnAddBackward0)
tensor(227.3427, grad_fnAddBackward0) 9.4 神经网络学习率优化
import torch
import torchvision
from torch import nn
from torch.nn import Conv2d, MaxPool2d, Flatten, Linear, Sequential
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriterdataset torchvision.datasets.CIFAR10(./dataset,trainFalse,transformtorchvision.transforms.ToTensor(),downloadTrue)
dataloader DataLoader(dataset, batch_size64,drop_lastTrue)class Tudui(nn.Module):def __init__(self):super(Tudui, self).__init__() self.model1 Sequential(Conv2d(3,32,5,padding2),MaxPool2d(2),Conv2d(32,32,5,padding2),MaxPool2d(2),Conv2d(32,64,5,padding2),MaxPool2d(2),Flatten(),Linear(1024,64),Linear(64,10))def forward(self, x):x self.model1(x)return xloss nn.CrossEntropyLoss() # 交叉熵
tudui Tudui()
optim torch.optim.SGD(tudui.parameters(),lr0.01) # 随机梯度下降优化器
scheduler torch.optim.lr_scheduler.StepLR(optim, step_size5, gamma0.1) # 每过 step_size 更新一次优化器更新是学习率为原来的学习率的的 0.1 倍
for epoch in range(20):running_loss 0.0for data in dataloader:imgs, targets dataoutputs tudui(imgs)result_loss loss(outputs, targets) # 计算实际输出与目标输出的差距optim.zero_grad() # 梯度清零result_loss.backward() # 反向传播计算损失函数的梯度optim.step() # 根据梯度对网络的参数进行调优scheduler.step() # 学习率太小了所以20个轮次后相当于没走多少running_loss running_loss result_lossprint(running_loss) # 对这一轮所有误差的总和
结果 Files already downloaded and verified
tensor(359.4722, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
tensor(359.4630, grad_fnAddBackward0)
