一.数据集简介
在动手撸MNIST之前相信大家对MNIST数据集也比较熟悉了，这里我也不多赘述了，只把关键的几个参数罗列一下：
类别：数字0~9 共计10个类别；
数量：总共7万张灰度图像，其中6万张训练，1万张测试，并且每张图都有标签，数字0对应标签0，数字1对应标签1，依次类推......
像素大小：28×28
通道数：单通道
二.网络简介
这里是一个比较简单的网络，其介绍大多在各论坛和论文期刊中可见，这里同样也不多赘述了
定义网络类：
import torch.nn as nn class Net(nn.Module): def __init__(self, in_c=784, out_c=10): super(Net, self).__init__() # 定义全连接层 self.fc1 = nn.Linear(in_c, 512) # 定义激活层 self.act1 = nn.ReLU(inplace=True) self.fc2 = nn.Linear(512, 256) self.act2 = nn.ReLU(inplace=True) self.fc3 = nn.Linear(256, 128) self.act3 = nn.ReLU(inplace=True) self.fc4 = nn.Linear(128, out_c) def forward(self, x): x = self.act1(self.fc1(x)) x = self.act2(self.fc2(x)) x = self.act3(self.fc3(x)) x = self.fc4(x) return x
三.动手撸CPU版本
网络具备，以此开始动收撸CPU版本的MNIST分类
流程
3.1 获取类（网络）
3.2 获取训练集和测试集
3.3 训练集和测试集载入（DataLoader）
3.4 定义损失函数--交叉熵函数
3.5 定义优化器--随机梯度下降
3.6 创建for循环A并创建空数组分别记录训练损失和精度
3.7 for循环A中：创建空数组分布记录测试损失和精度（注意此处应与训练分开）
3.8 设置训练次数
3.9 构建for循环B并在for循环中设置训练精度和损失值
3.10 在for循环B中：网络开始训练载入
3.11 在for循环B中：构建batch 和图像以及标签的for循环
3.12 在for循环B中：图像和标签Variable处理
3.13 在for循环B中：网络向前传播out = net(img) loss = criterion(out, label)
3.14 在for循环B中：记录误差
3.15 在for循环B中：计算分类正确率
3.16 结束B循环：此时还在A循环中，append损失和append精度
3.17 在A循环中新建循环此时的为测试集（测试集不训练）其他不变
CPU版本的demo:
import time import torch.nn as nn from torch import optim from torch.autograd import Variable from torch.utils.data import DataLoader from torchvision.datasets import mnist from torchvision import transforms class Net(nn.Module): def __init__(self, in_c=784, out_c=10): super(Net, self).__init__() # 定义全连接层 self.fc1 = nn.Linear(in_c, 512) # 定义激活层 self.act1 = nn.ReLU(inplace=True) self.fc2 = nn.Linear(512, 256) self.act2 = nn.ReLU(inplace=True) self.fc3 = nn.Linear(256, 128) self.act3 = nn.ReLU(inplace=True) self.fc4 = nn.Linear(128, out_c) def forward(self, x): x = self.act1(self.fc1(x)) x = self.act2(self.fc2(x)) x = self.act3(self.fc3(x)) x = self.fc4(x) return x t1 = time.time() # 搭建网络 net = Net() # 训练集 train_set = mnist.MNIST('./data', train=True, transform=transforms.ToTensor(), download=False) # 测试集 test_set = mnist.MNIST('./data', train=False, transform=transforms.ToTensor(), download=False) # 训练集载入器 train_data = DataLoader(train_set, batch_size=64, shuffle=True) # 测试集载入器 test_data = DataLoader(test_set, batch_size=64, shuffle=True) # 定义损失函数 -- 交叉熵 criterion = nn.CrossEntropyLoss() # 定义优化器 -- 随机梯度下降 optimizer = optim.SGD(net.parameters(), lr=0.01, weight_decay=0.00005) # 开始训练 losses = [] # 记录训练损失 acces = [] # 记录训练精度 eval_losses = [] # 记录测试损失 eval_acces = [] # 记录测试精度 nums_epoch = 20 # 训练次数 for epoch in range(nums_epoch): train_loss = 0 # 设置训练损失的初始值 train_acc = 0 # 设置训练精度的初始值 net.train() for batch, (img, label) in enumerate(train_data): img = img.reshape(img.size(0), -1) img = Variable(img) label = Variable(label) # 向前传播 out = net(img) loss = criterion(out, label) # 反向传播 optimizer.zero_grad() loss.backward() optimizer.step() # 记录误差 train_loss +=loss.item() # 计算分类正确率 _, pred = out.max(1) num_correct = (pred == label).sum().item() acc = num_correct / img.shape[0] if (batch +1) % 200 == 0: print('[INFO] Epoch-{}-Batch-{}: Train: Loss-{:.4f},Accuracy-{:.4f}'.format(epoch+1, batch+1, loss.item(),acc)) train_acc += acc losses.append(train_acc / len(train_data)) acces.append(train_acc / len(train_data)) eval_loss = 0 eval_acc = 0 # 测试集不训练 for img, label in test_data: img = img.reshape(img.size(0),-1) img = Variable(img) label = Variable(label) out = net(img) loss = criterion(out, label) eval_loss += loss.item() _, pred = out.max(1) num_correct = (pred == label).sum().item() acc = num_correct / img.shape[0] eval_acc += acc eval_losses.append(eval_loss / len(test_data)) eval_acces.append(eval_acc / len(test_data)) # 打印参数 set_epoch = epoch+1 set_lossTrain = train_loss / len(train_data) set_AccTrain = train_acc / len(train_data) set_lossEval = eval_loss / len(test_data) set_AccEval = eval_acc / len(test_data) print('[INFO] Epoch-{}: Train: Loss-{:.4f},Accuracy-{:.4f} |Test:Loss-{:.4f}, Accuracy-{:.4f}'.format(set_epoch, set_lossTrain, set_AccTrain, set_lossEval, set_AccEval)) t2 = time.time() t = t2 - t1 print(t)
CPU版本的demo(输出)：
[INFO] Epoch-1-Batch-200: Train: Loss-2.2869,Accuracy-0.2500 [INFO] Epoch-1-Batch-400: Train: Loss-2.2628,Accuracy-0.4062 [INFO] Epoch-1-Batch-600: Train: Loss-2.2056,Accuracy-0.5156 [INFO] Epoch-1-Batch-800: Train: Loss-1.9502,Accuracy-0.6875 [INFO] Epoch-1: Train: Loss-2.1569,Accuracy-0.0020 |Test:Loss-1.5634, Accuracy-0.6395 [INFO] Epoch-2-Batch-200: Train: Loss-0.9539,Accuracy-0.8125 [INFO] Epoch-2-Batch-400: Train: Loss-0.8835,Accuracy-0.7500 [INFO] Epoch-2-Batch-600: Train: Loss-0.5718,Accuracy-0.8281 [INFO] Epoch-2-Batch-800: Train: Loss-0.5851,Accuracy-0.8125 [INFO] Epoch-2: Train: Loss-0.8250,Accuracy-0.0034 |Test:Loss-0.5095, Accuracy-0.8504 [INFO] Epoch-3-Batch-200: Train: Loss-0.4938,Accuracy-0.7500 [INFO] Epoch-3-Batch-400: Train: Loss-0.5644,Accuracy-0.8438 [INFO] Epoch-3-Batch-600: Train: Loss-0.4656,Accuracy-0.8750 [INFO] Epoch-3-Batch-800: Train: Loss-0.4800,Accuracy-0.8438 [INFO] Epoch-3: Train: Loss-0.4432,Accuracy-0.0035 |Test:Loss-0.3720, Accuracy-0.8914 [INFO] Epoch-4-Batch-200: Train: Loss-0.3568,Accuracy-0.8750 [INFO] Epoch-4-Batch-400: Train: Loss-0.3659,Accuracy-0.8594 [INFO] Epoch-4-Batch-600: Train: Loss-0.3843,Accuracy-0.8281 [INFO] Epoch-4-Batch-800: Train: Loss-0.3291,Accuracy-0.8906 [INFO] Epoch-4: Train: Loss-0.3600,Accuracy-0.0037 |Test:Loss-0.3328, Accuracy-0.9015 [INFO] Epoch-5-Batch-200: Train: Loss-0.2843,Accuracy-0.8906 [INFO] Epoch-5-Batch-400: Train: Loss-0.2729,Accuracy-0.9375 [INFO] Epoch-5-Batch-600: Train: Loss-0.2628,Accuracy-0.9219 [INFO] Epoch-5-Batch-800: Train: Loss-0.1479,Accuracy-0.9531 [INFO] Epoch-5: Train: Loss-0.3174,Accuracy-0.0039 |Test:Loss-0.2917, Accuracy-0.9161 [INFO] Epoch-6-Batch-200: Train: Loss-0.3273,Accuracy-0.9062 [INFO] Epoch-6-Batch-400: Train: Loss-0.2906,Accuracy-0.9375 [INFO] Epoch-6-Batch-600: Train: Loss-0.2957,Accuracy-0.9062 [INFO] Epoch-6-Batch-800: Train: Loss-0.2804,Accuracy-0.9375 [INFO] Epoch-6: Train: Loss-0.2839,Accuracy-0.0039 |Test:Loss-0.2652, Accuracy-0.9247 [INFO] Epoch-7-Batch-200: Train: Loss-0.3675,Accuracy-0.8906 [INFO] Epoch-7-Batch-400: Train: Loss-0.3041,Accuracy-0.8906 [INFO] Epoch-7-Batch-600: Train: Loss-0.2421,Accuracy-0.9375 [INFO] Epoch-7-Batch-800: Train: Loss-0.1761,Accuracy-0.9219 [INFO] Epoch-7: Train: Loss-0.2561,Accuracy-0.0039 |Test:Loss-0.2401, Accuracy-0.9319 [INFO] Epoch-8-Batch-200: Train: Loss-0.1390,Accuracy-0.9531 [INFO] Epoch-8-Batch-400: Train: Loss-0.1204,Accuracy-0.9688 [INFO] Epoch-8-Batch-600: Train: Loss-0.1118,Accuracy-0.9844 [INFO] Epoch-8-Batch-800: Train: Loss-0.1276,Accuracy-0.9844 [INFO] Epoch-8: Train: Loss-0.2306,Accuracy-0.0041 |Test:Loss-0.2178, Accuracy-0.9365 [INFO] Epoch-9-Batch-200: Train: Loss-0.4543,Accuracy-0.9062 [INFO] Epoch-9-Batch-400: Train: Loss-0.3267,Accuracy-0.9219 [INFO] Epoch-9-Batch-600: Train: Loss-0.1870,Accuracy-0.9531 [INFO] Epoch-9-Batch-800: Train: Loss-0.3354,Accuracy-0.9062 [INFO] Epoch-9: Train: Loss-0.2094,Accuracy-0.0039 |Test:Loss-0.2016, Accuracy-0.9412 [INFO] Epoch-10-Batch-200: Train: Loss-0.1400,Accuracy-0.9219 [INFO] Epoch-10-Batch-400: Train: Loss-0.2871,Accuracy-0.9219 [INFO] Epoch-10-Batch-600: Train: Loss-0.1343,Accuracy-0.9531 [INFO] Epoch-10-Batch-800: Train: Loss-0.2881,Accuracy-0.8906 [INFO] Epoch-10: Train: Loss-0.1906,Accuracy-0.0039 |Test:Loss-0.1805, Accuracy-0.9460 [INFO] Epoch-11-Batch-200: Train: Loss-0.2244,Accuracy-0.9688 [INFO] Epoch-11-Batch-400: Train: Loss-0.1173,Accuracy-0.9688 [INFO] Epoch-11-Batch-600: Train: Loss-0.1551,Accuracy-0.9531 [INFO] Epoch-11-Batch-800: Train: Loss-0.1560,Accuracy-0.9531 [INFO] Epoch-11: Train: Loss-0.1748,Accuracy-0.0041 |Test:Loss-0.1693, Accuracy-0.9504 [INFO] Epoch-12-Batch-200: Train: Loss-0.2438,Accuracy-0.9688 [INFO] Epoch-12-Batch-400: Train: Loss-0.0888,Accuracy-0.9688 [INFO] Epoch-12-Batch-600: Train: Loss-0.0938,Accuracy-0.9688 [INFO] Epoch-12-Batch-800: Train: Loss-0.1019,Accuracy-0.9688 [INFO] Epoch-12: Train: Loss-0.1611,Accuracy-0.0041 |Test:Loss-0.1562, Accuracy-0.9515 [INFO] Epoch-13-Batch-200: Train: Loss-0.2955,Accuracy-0.9219 [INFO] Epoch-13-Batch-400: Train: Loss-0.3402,Accuracy-0.9062 [INFO] Epoch-13-Batch-600: Train: Loss-0.1040,Accuracy-0.9688 [INFO] Epoch-13-Batch-800: Train: Loss-0.1147,Accuracy-0.9844 [INFO] Epoch-13: Train: Loss-0.1491,Accuracy-0.0040 |Test:Loss-0.1475, Accuracy-0.9562 [INFO] Epoch-14-Batch-200: Train: Loss-0.0578,Accuracy-1.0000 [INFO] Epoch-14-Batch-400: Train: Loss-0.0836,Accuracy-0.9688 [INFO] Epoch-14-Batch-600: Train: Loss-0.1362,Accuracy-0.9688 [INFO] Epoch-14-Batch-800: Train: Loss-0.0897,Accuracy-0.9531 [INFO] Epoch-14: Train: Loss-0.1387,Accuracy-0.0041 |Test:Loss-0.1441, Accuracy-0.9561 [INFO] Epoch-15-Batch-200: Train: Loss-0.1424,Accuracy-0.9844 [INFO] Epoch-15-Batch-400: Train: Loss-0.0657,Accuracy-0.9844 [INFO] Epoch-15-Batch-600: Train: Loss-0.0836,Accuracy-0.9688 [INFO] Epoch-15-Batch-800: Train: Loss-0.1404,Accuracy-0.9688 [INFO] Epoch-15: Train: Loss-0.1289,Accuracy-0.0042 |Test:Loss-0.1301, Accuracy-0.9608 [INFO] Epoch-16-Batch-200: Train: Loss-0.1637,Accuracy-0.9219 [INFO] Epoch-16-Batch-400: Train: Loss-0.0509,Accuracy-1.0000 [INFO] Epoch-16-Batch-600: Train: Loss-0.2507,Accuracy-0.9375 [INFO] Epoch-16-Batch-800: Train: Loss-0.0801,Accuracy-0.9688 [INFO] Epoch-16: Train: Loss-0.1205,Accuracy-0.0041 |Test:Loss-0.1252, Accuracy-0.9610 [INFO] Epoch-17-Batch-200: Train: Loss-0.0761,Accuracy-0.9688 [INFO] Epoch-17-Batch-400: Train: Loss-0.0439,Accuracy-1.0000 [INFO] Epoch-17-Batch-600: Train: Loss-0.2204,Accuracy-0.9062 [INFO] Epoch-17-Batch-800: Train: Loss-0.0640,Accuracy-0.9844 [INFO] Epoch-17: Train: Loss-0.1128,Accuracy-0.0041 |Test:Loss-0.1211, Accuracy-0.9617 [INFO] Epoch-18-Batch-200: Train: Loss-0.0907,Accuracy-0.9844 [INFO] Epoch-18-Batch-400: Train: Loss-0.0587,Accuracy-0.9844 [INFO] Epoch-18-Batch-600: Train: Loss-0.0478,Accuracy-1.0000 [INFO] Epoch-18-Batch-800: Train: Loss-0.0532,Accuracy-0.9844 [INFO] Epoch-18: Train: Loss-0.1057,Accuracy-0.0042 |Test:Loss-0.1113, Accuracy-0.9654 [INFO] Epoch-19-Batch-200: Train: Loss-0.1051,Accuracy-0.9531 [INFO] Epoch-19-Batch-400: Train: Loss-0.1953,Accuracy-0.9219 [INFO] Epoch-19-Batch-600: Train: Loss-0.1334,Accuracy-0.9531 [INFO] Epoch-19-Batch-800: Train: Loss-0.1170,Accuracy-0.9531 [INFO] Epoch-19: Train: Loss-0.0991,Accuracy-0.0040 |Test:Loss-0.1087, Accuracy-0.9662 [INFO] Epoch-20-Batch-200: Train: Loss-0.0581,Accuracy-1.0000 [INFO] Epoch-20-Batch-400: Train: Loss-0.0779,Accuracy-0.9531 [INFO] Epoch-20-Batch-600: Train: Loss-0.1448,Accuracy-0.9375 [INFO] Epoch-20-Batch-800: Train: Loss-0.0859,Accuracy-0.9688 [INFO] Epoch-20: Train: Loss-0.0934,Accuracy-0.0041 |Test:Loss-0.1075, Accuracy-0.9661 143.9970304965973
四.CPU2GPU改进措施
在进行CPU2GPU的过程中基本流程不变，这里需要借助:
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')此语句的含义为当计算机存在cuda时则用GPU，若无GPU时则用CPU。
4.1注意事项
4.1.1. 灵活使用.to(device)可以使网络运行的更为顺畅并且batch也可以更大，更加合理分配资源
4.1.2. 这里需要用到cudnn,因此需要调用：import torch.backends.cudnn as cudnn
4.1.3. 分别对网络、数据、标签、损失函数 进行to(device)
4.1.4. 合理分配更多的batch
4.2具体可如下图所示：
4.2.1对网络进行to(device)和使用cudnn
4.2.2合理分配更多的batch
4.2.3 使用到torch.nn
4.2.4 分别对训练和测试的img label 进行to(device)
GPU版本的demo
import time import torch import torch.nn as nn import torch.backends.cudnn as cudnn from torch import optim from torch.autograd import Variable from torch.utils.data import DataLoader from torchvision.datasets import mnist from torchvision import transforms class Net(nn.Module): def __init__(self, in_c=784, out_c=10): super(Net, self).__init__() # 定义全连接层 self.fc1 = nn.Linear(in_c, 512) # 定义激活层 self.act1 = nn.ReLU(inplace=True) self.fc2 = nn.Linear(512, 256) self.act2 = nn.ReLU(inplace=True) self.fc3 = nn.Linear(256, 128) self.act3 = nn.ReLU(inplace=True) self.fc4 = nn.Linear(128, out_c) def forward(self, x): x = self.act1(self.fc1(x)) x = self.act2(self.fc2(x)) x = self.act3(self.fc3(x)) x = self.fc4(x) return x t1 = time.time() # 搭建网络 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') net = Net() cudnn.benchmark = True net = net.to(device) # 训练集 train_set = mnist.MNIST('./data', train=True, transform=transforms.ToTensor(), download=False) # 测试集 test_set = mnist.MNIST('./data', train=False, transform=transforms.ToTensor(), download=False) # 训练集载入器 train_data = DataLoader(train_set, batch_size=640, shuffle=True) # 测试集载入器 test_data = DataLoader(test_set, batch_size=320, shuffle=True) # 定义损失函数 -- 交叉熵 criterion = torch.nn.CrossEntropyLoss().to(device) # 定义优化器 -- 随机梯度下降 optimizer = optim.SGD(net.parameters(), lr=0.01, weight_decay=0.00005) # 开始训练 losses = [] # 记录训练损失 acces = [] # 记录训练精度 eval_losses = [] # 记录测试损失 eval_acces = [] # 记录测试精度 nums_epoch = 20 # 训练次数 for epoch in range(nums_epoch): train_loss = 0 # 设置训练损失的初始值 train_acc = 0 # 设置训练精度的初始值 net.train() for batch, (img, label) in enumerate(train_data): img = img.reshape(img.size(0), -1) img = Variable(img) img = img.to(device) label = Variable(label) label = label.to(device) # 向前传播 out = net(img) loss = criterion(out, label) # 反向传播 optimizer.zero_grad() loss.backward() optimizer.step() # 记录误差 train_loss += loss.item() # 计算分类正确率 _, pred = out.max(1) num_correct = (pred == label).sum().item() acc = num_correct / img.shape[0] if (batch + 1) % 200 == 0: print( '[INFO] Epoch-{}-Batch-{}: Train: Loss-{:.4f},Accuracy-{:.4f}'.format(epoch + 1, batch + 1, loss.item(), acc)) train_acc += acc losses.append(train_acc / len(train_data)) acces.append(train_acc / len(train_data)) eval_loss = 0 eval_acc = 0 # 测试集不训练 for img, label in test_data: img = img.reshape(img.size(0), -1) img = Variable(img) img = img.to(device) label = Variable(label) label = label.to(device) out = net(img) loss = criterion(out, label.to(device)) eval_loss += loss.item() _, pred = out.max(1) num_correct = (pred == label).sum().item() acc = num_correct / img.shape[0] eval_acc += acc eval_losses.append(eval_loss / len(test_data)) eval_acces.append(eval_acc / len(test_data)) # 打印参数 set_epoch = epoch + 1 set_lossTrain = train_loss / len(train_data) set_AccTrain = train_acc / len(train_data) set_lossEval = eval_loss / len(test_data) set_AccEval = eval_acc / len(test_data) print('[INFO] Epoch-{}: Train: Loss-{:.4f},Accuracy-{:.4f} |Test:Loss-{:.4f}, Accuracy-{:.4f}'.format(set_epoch, set_lossTrain, set_AccTrain, set_lossEval, set_AccEval)) t2 = time.time() t = t2 - t1 print(t)
GPU版本输出
[INFO] Epoch-1: Train: Loss-2.3025,Accuracy-0.0000 |Test:Loss-2.2982, Accuracy-0.1018 [INFO] Epoch-2: Train: Loss-2.2950,Accuracy-0.0000 |Test:Loss-2.2902, Accuracy-0.1034 [INFO] Epoch-3: Train: Loss-2.2867,Accuracy-0.0000 |Test:Loss-2.2812, Accuracy-0.1491 [INFO] Epoch-4: Train: Loss-2.2762,Accuracy-0.0000 |Test:Loss-2.2685, Accuracy-0.3172 [INFO] Epoch-5: Train: Loss-2.2611,Accuracy-0.0000 |Test:Loss-2.2498, Accuracy-0.4545 [INFO] Epoch-6: Train: Loss-2.2383,Accuracy-0.0000 |Test:Loss-2.2205, Accuracy-0.5208 [INFO] Epoch-7: Train: Loss-2.2020,Accuracy-0.0000 |Test:Loss-2.1736, Accuracy-0.5357 [INFO] Epoch-8: Train: Loss-2.1407,Accuracy-0.0000 |Test:Loss-2.0909, Accuracy-0.5215 [INFO] Epoch-9: Train: Loss-2.0353,Accuracy-0.0000 |Test:Loss-1.9529, Accuracy-0.5164 [INFO] Epoch-10: Train: Loss-1.8638,Accuracy-0.0000 |Test:Loss-1.7380, Accuracy-0.5524 [INFO] Epoch-11: Train: Loss-1.6231,Accuracy-0.0000 |Test:Loss-1.4687, Accuracy-0.6281 [INFO] Epoch-12: Train: Loss-1.3598,Accuracy-0.0000 |Test:Loss-1.2147, Accuracy-0.7030 [INFO] Epoch-13: Train: Loss-1.1373,Accuracy-0.0000 |Test:Loss-1.0222, Accuracy-0.7411 [INFO] Epoch-14: Train: Loss-0.9642,Accuracy-0.0000 |Test:Loss-0.8700, Accuracy-0.7768 [INFO] Epoch-15: Train: Loss-0.8361,Accuracy-0.0000 |Test:Loss-0.7663, Accuracy-0.7890 [INFO] Epoch-16: Train: Loss-0.7451,Accuracy-0.0000 |Test:Loss-0.6880, Accuracy-0.8087 [INFO] Epoch-17: Train: Loss-0.6791,Accuracy-0.0000 |Test:Loss-0.6312, Accuracy-0.8223 [INFO] Epoch-18: Train: Loss-0.6299,Accuracy-0.0000 |Test:Loss-0.5862, Accuracy-0.8302 [INFO] Epoch-19: Train: Loss-0.5907,Accuracy-0.0000 |Test:Loss-0.5545, Accuracy-0.8388 [INFO] Epoch-20: Train: Loss-0.5588,Accuracy-0.0000 |Test:Loss-0.5313, Accuracy-0.8452 84.11509966850281
自己敲的应该还不太对，还望指正
