1 深度学习步骤
（1）数据预处理：通过专门的数据加载，通过批训练提高模型表现，每次训练读取固定数量的样本输入到模型中进行训练
（2）深度神经网络搭建：逐层搭建，实现特定功能的层（如积层、池化层、批正则化层、LSTM层等）
（3）损失函数和优化器的设定：保证反向传播能够在用户定义的模型结构上实现
（4）模型训练：使用并行计算加速训练，将数据按批加载，放入GPU中训练，对损失函数反向传播回网络最前面的层，同时使用优化器调整网络参数
2 基本配置
导入相关的包
import os import numpy as py import torch import torch.nn as nn from torch.utils.data import Dataset, DataLoader import torch.optim as optimizer 统一设置超参数：batch size、初始学习率、训练次数、GPU配置 # set batch size batch_size = 16 # 初始学习率 lr = 1e-4 # 训练次数 max_epochs = 100 # 配置GPU device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu") device device(type='cuda', index=1)
3 数据读入
读取方式：通过Dataset+DataLoader的方式加载数据，Dataset定义好数据的格式和数据变换形式，DataLoader用iterative的方式不断读入批次数据。
自定义Dataset类：实现__init___、__getitem__、__len__函数
torch.utils.data.DataLoader参数：
batch_size：样本是按“批”读入的，表示每次读入的样本数
num_workers：表示用于读取数据的进程数
shuffle：是否将读入的数据打乱
drop_last：对于样本最后一部分没有达到批次数的样本，使其不再参与训练
4 模型构建
4.1 神经网络的构造
通过Module类构造模型，实例化模型之后，可完成模型构造
# 构造多层感知机 class MLP(nn.Module): def __init__(self, **kwargs): super(MLP, self).__init__(**kwargs) self.hidden = nn.Linear(784, 256) self.act = nn.ReLU() self.output = nn.Linear(256, 10) def forward(self, X): o = self.act(self.hidden(x)) return self.output(o) x = torch.rand(2, 784) net = MLP() print(x) net(x) tensor([[0.8924, 0.9624, 0.3262, ..., 0.8376, 0.1889, 0.9060], [0.3609, 0.8005, 0.5175, ..., 0.6255, 0.1462, 0.9846]]) tensor([[-0.0902, 0.0199, 0.0677, -0.0679, 0.0799, 0.0826, 0.0628, 0.1809, -0.2387, 0.0366], [-0.2271, 0.0056, -0.0984, -0.0432, -0.0160, -0.0038, 0.0953, 0.0545, -0.1530, -0.0214]], grad_fn=<AddmmBackward>)
4.2 神经网络常见的层
不含模型参数的层
# 构造一个输入减去均值后输出的层 class MyLayer(nn.Module): def __init__(self, **kwargs): super(MyLayer, self).__init__(**kwargs) def forward(self, x): return x - x.mean() x = torch.tensor([0, 5, 10, 15, 20], dtype=torch.float) layer = MyLayer() layer(x) tensor([-10., -5., 0., 5., 10.])
含模型参数的层：如果一个Tensor是Parameter，那么它会⾃动被添加到模型的参数列表里
# 使用ParameterList定义参数的列表 class MyListDense(nn.Module): def __init__(self): super(MyListDense, self).__init__() self.params = nn.ParameterList( [nn.Parameter(torch.randn(4, 4)) for i in range(3)]) self.params.append(nn.Parameter(torch.randn(4, 1))) def forward(self, x): for i in range(len(self.params)): x = torch.mm(x, self.params[i]) return x net = MyListDense() print(net) MyListDense( (params): ParameterList( (0): Parameter containing: [torch.FloatTensor of size 4x4] (1): Parameter containing: [torch.FloatTensor of size 4x4] (2): Parameter containing: [torch.FloatTensor of size 4x4] (3): Parameter containing: [torch.FloatTensor of size 4x1] ) ) # 使用ParameterDict定义参数的字典 class MyDictDense(nn.Module): def __init__(self): super(MyDictDense, self).__init__() self.params = nn.ParameterDict({ 'linear1': nn.Parameter(torch.randn(4, 4)), 'linear2': nn.Parameter(torch.randn(4, 1)) }) # 新增参数linear3 self.params.update({'linear3': nn.Parameter(torch.randn(4, 2))}) def forward(self, x, choice='linear1'): return torch.mm(x, self.params[choice]) net = MyDictDense() print(net) MyDictDense( (params): ParameterDict( (linear1): Parameter containing: [torch.FloatTensor of size 4x4] (linear2): Parameter containing: [torch.FloatTensor of size 4x1] (linear3): Parameter containing: [torch.FloatTensor of size 4x2] ) )
二维卷积层：使用nn.Conv2d类构造，模型参数包括卷积核和标量偏差，在训练模式时，通常先对卷积核随机初始化，再不断迭代卷积核和偏差
# 计算卷积层，对输入和输出做相应的升维和降维 def comp_conv2d(conv2d, X): # (1, 1)代表批量大小和通道数 X = X.view((1, 1) + X.shape) Y = conv2d(X) # 排除不关心的前两维：批量和通道 return Y.view(Y.shape[2:]) # 注意这里是两侧分别填充1⾏或列，所以在两侧一共填充2⾏或列 conv2d = nn.Conv2d(in_channels=1, out_channels=1, kernel_size=3,padding=1) X = torch.rand(8, 8) comp_conv2d(conv2d, X).shape torch.Size([8, 8]) 池化层：直接计算池化窗口内元素的最大值或者平均值，分别叫做最大池化或平均池化 # 二维池化层 def pool2d(X, pool_size, mode='max'): p_h, p_w = pool_size Y = torch.zeros((X.shape[0] - p_h + 1, X.shape[1] - p_w + 1)) for i in range(Y.shape[0]): for j in range(Y.shape[1]): if mode == 'max': Y[i, j] = X[i: i + p_h, j: j + p_w].max() elif mode == 'avg': Y[i, j] = X[i: i + p_h, j: j + p_w].mean() return Y X = torch.tensor([[0, 1, 2], [3, 4, 5], [6, 7, 8]], dtype=torch.float) pool2d(X, (2, 2), 'max') tensor([[4., 5.], [7., 8.]]) pool2d(X, (2, 2), 'avg') tensor([[2., 3.], [5., 6.]])
4.3 模型示例
神经网络训练过程：
定义可学习参数的神经网络
在输入数据集上进行迭代训练
通过神经网络处理输入数据
计算loss（损失值）
将梯度反向传播给神经网络参数
更新网络参数，使用梯度下降
LeNet(前馈神经网络)
import torch.nn.functional as F class Net(nn.Module): def __init__(self): super(Net, self).__init__() # 输入图像channel是1；输出channel是6；5x5卷积核 self.conv1 = nn.Conv2d(1, 6, 5) self.conv2 = nn.Conv2d(6, 16, 5) # an affine operation: y = Wx + b self.fc1 = nn.Linear(16 * 5 * 5, 120) self.fc2 = nn.Linear(120, 84) self.fc3 = nn.Linear(84, 10) def forward(self, x): # 2x2 Max pooling x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2)) # 如果是方阵,则可以只使用一个数字进行定义 x = F.max_pool2d(F.relu(self.conv2(x)), 2) x = x.view(-1, self.num_flat_features(x)) x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x = self.fc3(x) return x def num_flat_features(self, x): # 除去批处理维度的其他所有维度 size = x.size()[1:] num_features = 1 for s in size: num_features *= s return num_features net = Net() net Net( (conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1)) (conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1)) (fc1): Linear(in_features=400, out_features=120, bias=True) (fc2): Linear(in_features=120, out_features=84, bias=True) (fc3): Linear(in_features=84, out_features=10, bias=True) ) # 假设输入的数据为随机的32x32 input = torch.randn(1, 1, 32, 32) out = net(input) print(out) tensor([[-0.0921, -0.0605, -0.0726, -0.0451, 0.1399, -0.0087, 0.1075, 0.0799, -0.1472, 0.0288]], grad_fn=<AddmmBackward>) # 清零所有参数的梯度缓存，然后进行随机梯度的反向传播 net.zero_grad() out.backward(torch.randn(1, 10)) AlexNet class AlexNet(nn.Module): def __init__(self): super(AlexNet, self).__init__() self.conv = nn.Sequential( # in_channels, out_channels, kernel_size, stride, padding nn.Conv2d(1, 96, 11, 4), nn.ReLU(), # kernel_size, stride nn.MaxPool2d(3, 2), # 减小卷积窗口，使用填充为2来使得输入与输出的高和宽一致，且增大输出通道数 nn.Conv2d(96, 256, 5, 1, 2), nn.ReLU(), nn.MaxPool2d(3, 2), # 连续3个卷积层，且使用更小的卷积窗口。 # 除了最后的卷积层外，进一步增大了输出通道数。 # 前两个卷积层后不使用池化层来减小输入的高和宽 nn.Conv2d(256, 384, 3, 1, 1), nn.ReLU(), nn.Conv2d(384, 384, 3, 1, 1), nn.ReLU(), nn.Conv2d(384, 256, 3, 1, 1), nn.ReLU(), nn.MaxPool2d(3, 2) ) # 这里全连接层的输出个数比LeNet中的大数倍。使用丢弃层来缓解过拟合 self.fc = nn.Sequential( nn.Linear(256*5*5, 4096), nn.ReLU(), nn.Dropout(0.5), nn.Linear(4096, 4096), nn.ReLU(), nn.Dropout(0.5), # 输出层。由于这里使用Fashion-MNIST，所以用类别数为10，而非论文中的1000 nn.Linear(4096, 10), ) def forward(self, img): feature = self.conv(img) output = self.fc(feature.view(img.shape[0], -1)) return output net = AlexNet() print(net) AlexNet( (conv): Sequential( (0): Conv2d(1, 96, kernel_size=(11, 11), stride=(4, 4)) (1): ReLU() (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False) (3): Conv2d(96, 256, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2)) (4): ReLU() (5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False) (6): Conv2d(256, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (7): ReLU() (8): Conv2d(384, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (9): ReLU() (10): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (11): ReLU() (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False) ) (fc): Sequential( (0): Linear(in_features=6400, out_features=4096, bias=True) (1): ReLU() (2): Dropout(p=0.5, inplace=False) (3): Linear(in_features=4096, out_features=4096, bias=True) (4): ReLU() (5): Dropout(p=0.5, inplace=False) (6): Linear(in_features=4096, out_features=10, bias=True) ) )
5 损失函数
二分类交叉熵损失函数：torch.nn.BCELoss，用于计算二分类任务时的交叉熵
m = nn.Sigmoid() loss = nn.BCELoss() input = torch.randn(3, requires_grad=True) target = torch.empty(3).random_(2) output = loss(m(input), target) output.backward()
print('BCE损失函数的计算结果:',output)
BCE损失函数的计算结果: tensor(0.9389, grad_fn=<BinaryCrossEntropyBackward>)
交叉熵损失函数：torch.nn.CrossEntropyLoss，用于计算交叉熵
loss = nn.CrossEntropyLoss() input = torch.randn(3, 5, requires_grad=True) target = torch.empty(3, dtype=torch.long).random_(5) output = loss(input, target) output.backward()
print('CrossEntropy损失函数的计算结果:',output)
CrossEntropy损失函数的计算结果: tensor(2.7367, grad_fn=<NllLossBackward>)
L1损失函数：torch.nn.L1Loss，用于计算输出y和真实值target之差的绝对值
loss = nn.L1Loss() input = torch.randn(3, 5, requires_grad=True) target = torch.randn(3, 5) output = loss(input, target) output.backward()
print('L1损失函数的计算结果:',output)
L1损失函数的计算结果: tensor(1.0351, grad_fn=<L1LossBackward>)
MSE损失函数：torch.nn.MSELoss，用于计算输出y和真实值target之差的平方
loss = nn.MSELoss() input = torch.randn(3, 5, requires_grad=True) target = torch.randn(3, 5) output = loss(input, target) output.backward()
print('MSE损失函数的计算结果:',output)
MSE损失函数的计算结果: tensor(1.7612, grad_fn=<MseLossBackward>)
平滑L1（Smooth L1）损失函数：torch.nn.SmoothL1Loss，用于计算L1的平滑输出，减轻离群点带来的影响，通过与L1损失的比较，在0点的尖端处，过渡更为平滑
loss = nn.SmoothL1Loss() input = torch.randn(3, 5, requires_grad=True) target = torch.randn(3, 5) output = loss(input, target) output.backward()
print('Smooth L1损失函数的计算结果:',output)
Smooth L1损失函数的计算结果: tensor(0.7252, grad_fn=<SmoothL1LossBackward>)
目标泊松分布的负对数似然损失：torch.nn.PoissonNLLLoss
loss = nn.PoissonNLLLoss() log_input = torch.randn(5, 2, requires_grad=True) target = torch.randn(5, 2) output = loss(log_input, target) output.backward()
print('PoissonNL损失函数的计算结果:',output)
PoissonNL损失函数的计算结果: tensor(1.7593, grad_fn=<MeanBackward0>)
KL散度：torch.nn.KLDivLoss，用于连续分布的距离度量，可用在对离散采用的连续输出空间分布的回归场景
inputs = torch.tensor([[0.5, 0.3, 0.2], [0.2, 0.3, 0.5]]) target = torch.tensor([[0.9, 0.05, 0.05], [0.1, 0.7, 0.2]], dtype=torch.float) loss = nn.KLDivLoss(reduction='batchmean') output = loss(inputs,target)
print('KLDiv损失函数的计算结果:',output)
KLDiv损失函数的计算结果: tensor(-1.0006)
MarginRankingLoss：torch.nn.MarginRankingLoss，用于计算两组数据之间的差异（相似度），可使用在排序任务的场景
loss = nn.MarginRankingLoss() input1 = torch.randn(3, requires_grad=True) input2 = torch.randn(3, requires_grad=True) target = torch.randn(3).sign() output = loss(input1, input2, target) output.backward()
print('MarginRanking损失函数的计算结果:',output)
MarginRanking损失函数的计算结果: tensor(1.1762, grad_fn=<MeanBackward0>)
多标签边界损失函数：torch.nn.MultiLabelMarginLoss，用于计算多标签分类问题的损失
loss = nn.MultiLabelMarginLoss() x = torch.FloatTensor([[0.9, 0.2, 0.4, 0.8]]) # 真实的分类是，第3类和第0类 y = torch.LongTensor([[3, 0, -1, 1]]) output = loss(x, y) print('MultiLabelMargin损失函数的计算结果:',output) MultiLabelMargin损失函数的计算结果: tensor(0.4500)
二分类损失函数：torch.nn.SoftMarginLoss，用于计算二分类的logistic损失
# 定义两个样本，两个神经元
inputs = torch.tensor([[0.3, 0.7], [0.5, 0.5]])
target = torch.tensor([[-1, 1], [1, -1]], dtype=torch.float)
# 该loss对每个神经元计算，需要为每个神经元单独设置标签
loss_f = nn.SoftMarginLoss()
output = loss_f(inputs, target)
print('SoftMargin损失函数的计算结果:',output)
SoftMargin损失函数的计算结果: tensor(0.6764)
多分类的折页损失函数：torch.nn.MultiMarginLoss，用于计算多分类问题的折页损失
inputs = torch.tensor([[0.3, 0.7], [0.5, 0.5]])
target = torch.tensor([0, 1], dtype=torch.long)
loss_f = nn.MultiMarginLoss()
output = loss_f(inputs, target)
print('MultiMargin损失函数的计算结果:',output)
MultiMargin损失函数的计算结果: tensor(0.6000)
三元组损失函数：torch.nn.TripletMarginLoss，用于处理<实体1，关系，实体2>类型的数据，计算该类型数据的损失
triplet_loss = nn.TripletMarginLoss(margin=1.0, p=2)
anchor = torch.randn(100, 128, requires_grad=True)
positive = torch.randn(100, 128, requires_grad=True)
negative = torch.randn(100, 128, requires_grad=True)
output = triplet_loss(anchor, positive, negative)
output.backward()
print('TripletMargin损失函数的计算结果:',output)
TripletMargin损失函数的计算结果: tensor(1.1507, grad_fn=<MeanBackward0>)
HingEmbeddingLoss：torch.nn.HingeEmbeddingLoss，用于计算输出的embedding
结果的Hing损失
loss_f = nn.HingeEmbeddingLoss()
inputs = torch.tensor([[1., 0.8, 0.5]])
target = torch.tensor([[1, 1, -1]])
output = loss_f(inputs,target)
print('HingEmbedding损失函数的计算结果:',output)
HingEmbedding损失函数的计算结果: tensor(0.7667)
余弦相似度：torch.nn.CosineEmbeddingLoss，用于计算两个向量的余弦相似度，如果两个向量距离近，则损失函数值小，反之亦然
loss_f = nn.CosineEmbeddingLoss()
inputs_1 = torch.tensor([[0.3, 0.5, 0.7], [0.3, 0.5, 0.7]])
inputs_2 = torch.tensor([[0.1, 0.3, 0.5], [0.1, 0.3, 0.5]])
target = torch.tensor([1, -1], dtype=torch.float)
output = loss_f(inputs_1,inputs_2,target)
print('CosineEmbedding损失函数的计算结果:',output)
CosineEmbedding损失函数的计算结果: tensor(0.5000)
CTC损失函数：torch.nn.CTCLoss，用于处理时序数据的分类问题，计算连续时间序列和目标序列之间的损失
# Target are to be padded # 序列长度 T = 50 # 类别数（包括空类） C = 20 # batch size N = 16 # Target sequence length of longest target in batch (padding length) S = 30 # Minimum target length, for demonstration purposes S_min = 10 input = torch.randn(T, N, C).log_softmax(2).detach().requires_grad_() # 初始化target(0 = blank, 1:C = classes) target = torch.randint(low=1, high=C, size=(N, S), dtype=torch.long) input_lengths = torch.full(size=(N,), fill_value=T, dtype=torch.long) target_lengths = torch.randint(low=S_min, high=S, size=(N,), dtype=torch.long) ctc_loss = nn.CTCLoss() loss = ctc_loss(input, target, input_lengths, target_lengths) loss.backward()
print('CTC损失函数的计算结果:',loss)
CTC损失函数的计算结果: tensor(6.1333, grad_fn=<MeanBackward0>)
6 优化器
6.1 Optimizer的属性和方法
使用方向：为了使求解参数过程更快，使用BP+优化器逼近求解
Optimizer的属性：
defaults：优化器的超参数
state：参数的缓存
param_groups：参数组，顺序是params，lr，momentum，dampening，weight_decay，nesterov
Optimizer的方法：
zero_grad()：清空所管理参数的梯度
step()：执行一步梯度更新
add_param_group()：添加参数组
load_state_dict()：加载状态参数字典，可以用来进行模型的断点续训练，继续上次的参数进行训练
state_dict()：获取优化器当前状态信息字典
6.2 基本操作
# 设置权重，服从正态分布 --> 2 x 2 weight = torch.randn((2, 2), requires_grad=True) # 设置梯度为全1矩阵 --> 2 x 2 weight.grad = torch.ones((2, 2)) # 输出现有的weight和data print("The data of weight before step:\n{}".format(weight.data)) print("The grad of weight before step:\n{}".format(weight.grad)) The data of weight before step: tensor([[-0.5871, -1.1311], [-1.0446, 0.2656]]) The grad of weight before step: tensor([[1., 1.], [1., 1.]]) # 实例化优化器 optimizer = torch.optim.SGD([weight], lr=0.1, momentum=0.9) # 进行一步操作 optimizer.step() # 查看进行一步后的值，梯度 print("The data of weight after step:\n{}".format(weight.data)) print("The grad of weight after step:\n{}".format(weight.grad)) The data of weight after step: tensor([[-0.6871, -1.2311], [-1.1446, 0.1656]]) The grad of weight after step: tensor([[1., 1.], [1., 1.]]) # 权重清零 optimizer.zero_grad() # 检验权重是否为0 print("The grad of weight after optimizer.zero_grad():\n{}".format(weight.grad)) The grad of weight after optimizer.zero_grad(): tensor([[0., 0.], [0., 0.]]) # 添加参数：weight2 weight2 = torch.randn((3, 3), requires_grad=True) optimizer.add_param_group({"params": weight2, 'lr': 0.0001, 'nesterov': True}) # 查看现有的参数 print("optimizer.param_groups is\n{}".format(optimizer.param_groups)) # 查看当前状态信息 opt_state_dict = optimizer.state_dict() print("state_dict before step:\n", opt_state_dict) optimizer.param_groups is [{'params': [tensor([[-0.6871, -1.2311], [-1.1446, 0.1656]], requires_grad=True)], 'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False}, {'params': [tensor([[ 0.0411, -0.6569, 0.7445], [-0.7056, 1.1146, -0.4409], [-0.2302, -1.1507, -1.3807]], requires_grad=True)], 'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0}] state_dict before step: {'state': {0: {'momentum_buffer': tensor([[1., 1.], [1., 1.]])}}, 'param_groups': [{'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False, 'params': [0]}, {'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'params': [1]}]} # 进行5次step操作 for _ in range(50): optimizer.step() # 输出现有状态信息 print("state_dict after step:\n", optimizer.state_dict()) state_dict after step: {'state': {0: {'momentum_buffer': tensor([[0.0052, 0.0052], [0.0052, 0.0052]])}}, 'param_groups': [{'lr': 0.1, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'nesterov': False, 'params': [0]}, {'lr': 0.0001, 'nesterov': True, 'momentum': 0.9, 'dampening': 0, 'weight_decay': 0, 'params': [1]}]}
7 训练与评估
def train(epoch): # 设置训练状态 model.train() train_loss = 0 # 循环读取DataLoader中的全部数据 for data, label in train_loader: # 将数据放到GPU用于后续计算 data, label = data.cuda(), label.cuda() # 将优化器的梯度清0 optimizer.zero_grad() # 将数据输入给模型 output = model(data) # 设置损失函数 loss = criterion(label, output) # 将loss反向传播给网络 loss.backward() # 使用优化器更新模型参数 optimizer.step() # 累加训练损失 train_loss += loss.item()*data.size(0) train_loss = train_loss/len(train_loader.dataset) print('Epoch: {} \tTraining Loss: {:.6f}'.format(epoch, train_loss)) def val(epoch): # 设置验证状态 model.eval() val_loss = 0 # 不设置梯度 with torch.no_grad(): for data, label in val_loader: data, label = data.cuda(), label.cuda() output = model(data) preds = torch.argmax(output, 1) loss = criterion(output, label) val_loss += loss.item()*data.size(0) # 计算准确率 running_accu += torch.sum(preds == label.data) val_loss = val_loss/len(val_loader.dataset) print('Epoch: {} \tTraining Loss: {:.6f}'.format(epoch, val_loss))
