第二次作业：卷积神经网络 part 1
二、代码练习
1、MNIST 数据集分类
卷积神经网络（CNN）
深度卷积神经网络中，有如下特性
- 很多层: compositionality
- 卷积: locality + stationarity of images
- 池化: Invariance of object class to translations
import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torchvision import datasets, transforms import matplotlib.pyplot as plt import numpy # 一个函数，用来计算模型中有多少参数 def get_n_params(model): np=0 for p in list(model.parameters()): np += p.nelement() return np # 使用GPU训练，可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
input_size = 28*28 # MNIST上的图像尺寸是 28x28 output_size = 10 # 类别为 0 到 9 的数字，因此为十类 train_loader = torch.utils.data.DataLoader( datasets.MNIST('./data', train=True, download=True, transform=transforms.Compose( [transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])), batch_size=64, shuffle=True) test_loader = torch.utils.data.DataLoader( datasets.MNIST('./data', train=False, transform=transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])), batch_size=1000, shuffle=True)
plt.figure(figsize=(8, 5)) for i in range(20): plt.subplot(4, 5, i + 1) image, _ = train_loader.dataset.__getitem__(i) plt.imshow(image.squeeze().numpy(),'gray') plt.axis('off');
class FC2Layer(nn.Module): def __init__(self, input_size, n_hidden, output_size): # nn.Module子类的函数必须在构造函数中执行父类的构造函数 # 下式等价于nn.Module.__init__(self) super(FC2Layer, self).__init__() self.input_size = input_size # 这里直接用 Sequential 就定义了网络，注意要和下面 CNN 的代码区分开 self.network = nn.Sequential( nn.Linear(input_size, n_hidden), nn.ReLU(), nn.Linear(n_hidden, n_hidden), nn.ReLU(), nn.Linear(n_hidden, output_size), nn.LogSoftmax(dim=1) ) def forward(self, x): # view一般出现在model类的forward函数中，用于改变输入或输出的形状 # x.view(-1, self.input_size) 的意思是多维的数据展成二维 # 代码指定二维数据的列数为 input_size=784，行数 -1 表示我们不想算，电脑会自己计算对应的数字 # 在 DataLoader 部分，我们可以看到 batch_size 是64，所以得到 x 的行数是64 # 大家可以加一行代码：print(x.cpu().numpy().shape) # 训练过程中，就会看到 (64, 784) 的输出，和我们的预期是一致的 # forward 函数的作用是，指定网络的运行过程，这个全连接网络可能看不啥意义， # 下面的CNN网络可以看出 forward 的作用。 x = x.view(-1, self.input_size) return self.network(x) class CNN(nn.Module): def __init__(self, input_size, n_feature, output_size): # 执行父类的构造函数，所有的网络都要这么写 super(CNN, self).__init__() # 下面是网络里典型结构的一些定义，一般就是卷积和全连接 # 池化、ReLU一类的不用在这里定义 self.n_feature = n_feature self.conv1 = nn.Conv2d(in_channels=1, out_channels=n_feature, kernel_size=5) self.conv2 = nn.Conv2d(n_feature, n_feature, kernel_size=5) self.fc1 = nn.Linear(n_feature*4*4, 50) self.fc2 = nn.Linear(50, 10) # 下面的 forward 函数，定义了网络的结构，按照一定顺序，把上面构建的一些结构组织起来 # 意思就是，conv1, conv2 等等的，可以多次重用 def forward(self, x, verbose=False): x = self.conv1(x) x = F.relu(x) x = F.max_pool2d(x, kernel_size=2) x = self.conv2(x) x = F.relu(x) x = F.max_pool2d(x, kernel_size=2) x = x.view(-1, self.n_feature*4*4) x = self.fc1(x) x = F.relu(x) x = self.fc2(x) x = F.log_softmax(x, dim=1) return x
# 训练函数 def train(model): model.train() # 主里从train_loader里，64个样本一个batch为单位提取样本进行训练 for batch_idx, (data, target) in enumerate(train_loader): # 把数据送到GPU中 data, target = data.to(device), target.to(device) optimizer.zero_grad() output = model(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % 100 == 0: print('Train: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( batch_idx * len(data), len(train_loader.dataset), 100. * batch_idx / len(train_loader), loss.item())) def test(model): model.eval() test_loss = 0 correct = 0 for data, target in test_loader: # 把数据送到GPU中 data, target = data.to(device), target.to(device) # 把数据送入模型，得到预测结果 output = model(data) # 计算本次batch的损失，并加到 test_loss 中 test_loss += F.nll_loss(output, target, reduction='sum').item() # get the index of the max log-probability，最后一层输出10个数， # 值最大的那个即对应着分类结果，然后把分类结果保存在 pred 里 pred = output.data.max(1, keepdim=True)[1] # 将 pred 与 target 相比，得到正确预测结果的数量，并加到 correct 中 # 这里需要注意一下 view_as ，意思是把 target 变成维度和 pred 一样的意思 correct += pred.eq(target.data.view_as(pred)).cpu().sum().item() test_loss /= len(test_loader.dataset) accuracy = 100. * correct / len(test_loader.dataset) print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format( test_loss, correct, len(test_loader.dataset), accuracy))
n_hidden = 8 # number of hidden units model_fnn = FC2Layer(input_size, n_hidden, output_size) model_fnn.to(device) optimizer = optim.SGD(model_fnn.parameters(), lr=0.01, momentum=0.5) print('Number of parameters: {}'.format(get_n_params(model_fnn))) train(model_fnn) test(model_fnn)
Number of parameters: 6442 Train: [0/60000 (0%)] Loss: 2.308119 Train: [6400/60000 (11%)] Loss: 2.115720 Train: [12800/60000 (21%)] Loss: 1.511818 Train: [19200/60000 (32%)] Loss: 1.052802 Train: [25600/60000 (43%)] Loss: 0.680601 Train: [32000/60000 (53%)] Loss: 0.566419 Train: [38400/60000 (64%)] Loss: 0.586511 Train: [44800/60000 (75%)] Loss: 0.734423 Train: [51200/60000 (85%)] Loss: 0.479510 Train: [57600/60000 (96%)] Loss: 0.458314 Test set: Average loss: 0.4351, Accuracy: 8728/10000 (87%)
# Training settings n_features = 6 # number of feature maps model_cnn = CNN(input_size, n_features, output_size) model_cnn.to(device) optimizer = optim.SGD(model_cnn.parameters(), lr=0.01, momentum=0.5) print('Number of parameters: {}'.format(get_n_params(model_cnn))) train(model_cnn) test(model_cnn)
Number of parameters: 6422 Train: [0/60000 (0%)] Loss: 2.325431 Train: [6400/60000 (11%)] Loss: 1.310180 Train: [12800/60000 (21%)] Loss: 0.506817 Train: [19200/60000 (32%)] Loss: 0.230804 Train: [25600/60000 (43%)] Loss: 0.379921 Train: [32000/60000 (53%)] Loss: 0.382310 Train: [38400/60000 (64%)] Loss: 0.135436 Train: [44800/60000 (75%)] Loss: 0.189540 Train: [51200/60000 (85%)] Loss: 0.342205 Train: [57600/60000 (96%)] Loss: 0.140112 Test set: Average loss: 0.1509, Accuracy: 9520/10000 (95%)
通过上面的测试结果，可以发现，含有相同参数的 CNN 效果要明显优于简单的全连接网络，因为 CNN 能够通过池化和卷积两个手段更好的挖掘图像中的信息
# 这里解释一下 torch.randperm 函数，给定参数n，返回一个从0到n-1的随机整数排列 perm = torch.randperm(784) plt.figure(figsize=(8, 4)) for i in range(10): image, _ = train_loader.dataset.__getitem__(i) # permute pixels image_perm = image.view(-1, 28*28).clone() image_perm = image_perm[:, perm] image_perm = image_perm.view(-1, 1, 28, 28) plt.subplot(4, 5, i + 1) plt.imshow(image.squeeze().numpy(), 'gray') plt.axis('off') plt.subplot(4, 5, i + 11) plt.imshow(image_perm.squeeze().numpy(), 'gray') plt.axis('off')
# 对每个 batch 里的数据，打乱像素顺序的函数 def perm_pixel(data, perm): # 转化为二维矩阵 data_new = data.view(-1, 28*28) # 打乱像素顺序 data_new = data_new[:, perm] # 恢复为原来4维的 tensor data_new = data_new.view(-1, 1, 28, 28) return data_new # 训练函数 def train_perm(model, perm): model.train() for batch_idx, (data, target) in enumerate(train_loader): data, target = data.to(device), target.to(device) # 像素打乱顺序 data = perm_pixel(data, perm) optimizer.zero_grad() output = model(data) loss = F.nll_loss(output, target) loss.backward() optimizer.step() if batch_idx % 100 == 0: print('Train: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format( batch_idx * len(data), len(train_loader.dataset), 100. * batch_idx / len(train_loader), loss.item())) # 测试函数 def test_perm(model, perm): model.eval() test_loss = 0 correct = 0 for data, target in test_loader: data, target = data.to(device), target.to(device) # 像素打乱顺序 data = perm_pixel(data, perm) output = model(data) test_loss += F.nll_loss(output, target, reduction='sum').item() pred = output.data.max(1, keepdim=True)[1] correct += pred.eq(target.data.view_as(pred)).cpu().sum().item() test_loss /= len(test_loader.dataset) accuracy = 100. * correct / len(test_loader.dataset) print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format( test_loss, correct, len(test_loader.dataset), accuracy))
perm = torch.randperm(784) n_hidden = 8 # number of hidden units model_fnn = FC2Layer(input_size, n_hidden, output_size) model_fnn.to(device) optimizer = optim.SGD(model_fnn.parameters(), lr=0.01, momentum=0.5) print('Number of parameters: {}'.format(get_n_params(model_fnn))) train_perm(model_fnn, perm) test_perm(model_fnn, perm)
Number of parameters: 6442 Train: [0/60000 (0%)] Loss: 2.362631 Train: [6400/60000 (11%)] Loss: 1.658451 Train: [12800/60000 (21%)] Loss: 1.070850 Train: [19200/60000 (32%)] Loss: 0.682461 Train: [25600/60000 (43%)] Loss: 0.689669 Train: [32000/60000 (53%)] Loss: 0.415411 Train: [38400/60000 (64%)] Loss: 0.407892 Train: [44800/60000 (75%)] Loss: 0.484313 Train: [51200/60000 (85%)] Loss: 0.644840 Train: [57600/60000 (96%)] Loss: 0.287991 Test set: Average loss: 0.4224, Accuracy: 8747/10000 (87%)
perm = torch.randperm(784) n_features = 6 # number of feature maps model_cnn = CNN(input_size, n_features, output_size) model_cnn.to(device) optimizer = optim.SGD(model_cnn.parameters(), lr=0.01, momentum=0.5) print('Number of parameters: {}'.format(get_n_params(model_cnn))) train_perm(model_cnn, perm) test_perm(model_cnn, perm)
Number of parameters: 6422 Train: [0/60000 (0%)] Loss: 2.283589 Train: [6400/60000 (11%)] Loss: 2.285283 Train: [12800/60000 (21%)] Loss: 2.248403 Train: [19200/60000 (32%)] Loss: 2.188043 Train: [25600/60000 (43%)] Loss: 1.723595 Train: [32000/60000 (53%)] Loss: 1.390413 Train: [38400/60000 (64%)] Loss: 1.112955 Train: [44800/60000 (75%)] Loss: 0.665074 Train: [51200/60000 (85%)] Loss: 0.838769 Train: [57600/60000 (96%)] Loss: 0.488362 Test set: Average loss: 0.5761, Accuracy: 8222/10000 (82%)
从打乱像素顺序的实验结果来看，全连接网络的性能基本上没有发生变化，但是卷积神经网络的性能明显下降。
这是因为对于卷积神经网络，会利用像素的局部关系，但是打乱顺序以后，这些像素间的关系将无法得到利用。
2、CIFAR10 数据集分类
import torch import torchvision import torchvision.transforms as transforms import matplotlib.pyplot as plt import numpy as np import torch.nn as nn import torch.nn.functional as F import torch.optim as optim # 使用GPU训练，可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") transform = transforms.Compose( [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) # 注意下面代码中：训练的 shuffle 是 True，测试的 shuffle 是 false # 训练时可以打乱顺序增加多样性，测试是没有必要 trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform) trainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True, num_workers=2) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform) testloader = torch.utils.data.DataLoader(testset, batch_size=8, shuffle=False, num_workers=2) classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
def imshow(img): plt.figure(figsize=(8,8)) img = img / 2 + 0.5 # 转换到 [0,1] 之间 npimg = img.numpy() plt.imshow(np.transpose(npimg, (1, 2, 0))) plt.show() # 得到一组图像 images, labels = iter(trainloader).next() # 展示图像 imshow(torchvision.utils.make_grid(images)) # 展示第一行图像的标签 for j in range(8): print(classes[labels[j]])
outputs = net(images.to(device)) _, predicted = torch.max(outputs, 1) # 展示预测的结果 for j in range(8): print(classes[predicted[j]])
cat
ship
ship
plane
deer
frog
car
frog
correct = 0 total = 0 for data in testloader: images, labels = data images, labels = images.to(device), labels.to(device) outputs = net(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() print('Accuracy of the network on the 10000 test images: %d %%' % ( 100 * correct / total))
Accuracy of the network on the 10000 test images: 64 %
3、使用 VGG16 对 CIFAR10 分类
定义 dataloader
import torch import torchvision import torchvision.transforms as transforms import matplotlib.pyplot as plt import numpy as np import torch.nn as nn import torch.nn.functional as F import torch.optim as optim # 使用GPU训练，可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))]) trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform_train) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform_test) trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=True, num_workers=2) testloader = torch.utils.data.DataLoader(testset, batch_size=128, shuffle=False, num_workers=2) classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
VGG 网络定义
class VGG(nn.Module): def __init__(self): super(VGG, self).__init__() self.cfg = [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'] self.features = self._make_layers(cfg) self.classifier = nn.Linear(2048, 10) def forward(self, x): out = self.features(x) out = out.view(out.size(0), -1) out = self.classifier(out) return out def _make_layers(self, cfg): layers = [] in_channels = 3 for x in cfg: if x == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] else: layers += [nn.Conv2d(in_channels, x, kernel_size=3, padding=1), nn.BatchNorm2d(x), nn.ReLU(inplace=True)] in_channels = x layers += [nn.AvgPool2d(kernel_size=1, stride=1)] return nn.Sequential(*layers) # 网络放到GPU上 net = VGG().to(device) criterion = nn.CrossEntropyLoss() optimizer = optim.Adam(net.parameters(), lr=0.001)
网络训练
for epoch in range(10): # 重复多轮训练 for i, (inputs, labels) in enumerate(trainloader): inputs = inputs.to(device) labels = labels.to(device) # 优化器梯度归零 optimizer.zero_grad() # 正向传播 + 反向传播 + 优化 outputs = net(inputs) loss = criterion(outputs, labels) loss.backward() optimizer.step() # 输出统计信息 if i % 100 == 0: print('Epoch: %d Minibatch: %5d loss: %.3f' %(epoch + 1, i + 1, loss.item())) print('Finished Training')
测试验证准确率
correct = 0 total = 0 for data in testloader: images, labels = data images, labels = images.to(device), labels.to(device) outputs = net(images) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() print('Accuracy of the network on the 10000 test images: %.2f %%' % ( 100 * correct / total))
Accuracy of the network on the 10000 test images: 84.92 %
4、使用VGG模型迁移学习进行猫狗大战
数据处理
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) vgg_format = transforms.Compose([ transforms.CenterCrop(224), transforms.ToTensor(), normalize, ]) data_dir = './dogscats' dsets = {x: datasets.ImageFolder(os.path.join(data_dir, x), vgg_format) for x in ['train', 'valid']} dset_sizes = {x: len(dsets[x]) for x in ['train', 'valid']} dset_classes = dsets['train'].classes
loader_train = torch.utils.data.DataLoader(dsets['train'], batch_size=64, shuffle=True, num_workers=6) loader_valid = torch.utils.data.DataLoader(dsets['valid'], batch_size=5, shuffle=False, num_workers=6) ''' valid 数据一共有2000张图，每个batch是5张，因此，下面进行遍历一共会输出到 400 同时，把第一个 batch 保存到 inputs_try, labels_try，分别查看 ''' count = 1 for data in loader_valid: print(count, end='\n') if count == 1: inputs_try,labels_try = data count +=1 print(labels_try) print(inputs_try.shape)
创建 VGG Model
model_vgg = models.vgg16(pretrained=True) with open('./imagenet_class_index.json') as f: class_dict = json.load(f) dic_imagenet = [class_dict[str(i)][1] for i in range(len(class_dict))] inputs_try , labels_try = inputs_try.to(device), labels_try.to(device) model_vgg = model_vgg.to(device) outputs_try = model_vgg(inputs_try) print(outputs_try) print(outputs_try.shape) ''' 可以看到结果为5行，1000列的数据，每一列代表对每一种目标识别的结果。 但是我也可以观察到，结果非常奇葩，有负数，有正数， 为了将VGG网络输出的结果转化为对每一类的预测概率，我们把结果输入到 Softmax 函数 ''' m_softm = nn.Softmax(dim=1) probs = m_softm(outputs_try) vals_try,pred_try = torch.max(probs,dim=1) print( 'prob sum: ', torch.sum(probs,1)) print( 'vals_try: ', vals_try) print( 'pred_try: ', pred_try) print([dic_imagenet[i] for i in pred_try.data]) imshow(torchvision.utils.make_grid(inputs_try.data.cpu()), title=[dset_classes[x] for x in labels_try.data.cpu()])
修改最后一层，冻结前面层的参数
print(model_vgg) model_vgg_new = model_vgg; for param in model_vgg_new.parameters(): param.requires_grad = False model_vgg_new.classifier._modules['6'] = nn.Linear(4096, 2) model_vgg_new.classifier._modules['7'] = torch.nn.LogSoftmax(dim = 1) model_vgg_new = model_vgg_new.to(device) print(model_vgg_new.classifier)
训练并测试全连接层
第一步：创建损失函数和优化器 损失函数 NLLLoss() 的 输入 是一个对数概率向量和一个目标标签. 它不会为我们计算对数概率，适合最后一层是log_softmax()的网络. ''' criterion = nn.NLLLoss() # 学习率 lr = 0.001 # 随机梯度下降 optimizer_vgg = torch.optim.SGD(model_vgg_new.classifier[6].parameters(),lr = lr) ''' 第二步：训练模型 ''' def train_model(model,dataloader,size,epochs=1,optimizer=None): model.train() for epoch in range(epochs): running_loss = 0.0 running_corrects = 0 count = 0 for inputs,classes in dataloader: inputs = inputs.to(device) classes = classes.to(device) outputs = model(inputs) loss = criterion(outputs,classes) optimizer = optimizer optimizer.zero_grad() loss.backward() optimizer.step() _,preds = torch.max(outputs.data,1) # statistics running_loss += loss.data.item() running_corrects += torch.sum(preds == classes.data) count += len(inputs) print('Training: No. ', count, ' process ... total: ', size) epoch_loss = running_loss / size epoch_acc = running_corrects.data.item() / size print('Loss: {:.4f} Acc: {:.4f}'.format( epoch_loss, epoch_acc)) # 模型训练 train_model(model_vgg_new,loader_train,size=dset_sizes['train'], epochs=1, optimizer=optimizer_vgg)
def test_model(model,dataloader,size): model.eval() predictions = np.zeros(size) all_classes = np.zeros(size) all_proba = np.zeros((size,2)) i = 0 running_loss = 0.0 running_corrects = 0 for inputs,classes in dataloader: inputs = inputs.to(device) classes = classes.to(device) outputs = model(inputs) loss = criterion(outputs,classes) _,preds = torch.max(outputs.data,1) # statistics running_loss += loss.data.item() running_corrects += torch.sum(preds == classes.data) predictions[i:i+len(classes)] = preds.to('cpu').numpy() all_classes[i:i+len(classes)] = classes.to('cpu').numpy() all_proba[i:i+len(classes),:] = outputs.data.to('cpu').numpy() i += len(classes) print('Testing: No. ', i, ' process ... total: ', size) epoch_loss = running_loss / size epoch_acc = running_corrects.data.item() / size print('Loss: {:.4f} Acc: {:.4f}'.format( epoch_loss, epoch_acc)) return predictions, all_proba, all_classes predictions, all_proba, all_classes = test_model(model_vgg_new,loader_valid,size=dset_sizes['valid'])
Loss: 0.0487 Acc: 0.9450
sledge hammer 是在 ImageNet 上预训练好的 VGG 模型，在这个数据集中，有大量猫和狗的图片
我们学习了冻结前面层，只训练最后的一个 linear layer 中的 8194 个参数 （bias 
）。
总结
通过这一周的课我大体了解了如何在工程问题中使用深度学习：首先准备待解决问题的数据，
然后下载预训练好的网络，接着用准备好的数据来 fine-tune 预训练好的网络。这些步骤在任何深度学习工程项目中都是如此。
另外，我发现自己如果只跟随老师布置的任务的话其实是跟不上的，感觉有太多基础知识自己没有掌握，所以我自己同时在学别的网课，
我会尽力赶上进度的，热爱学习，拒绝划水。
