从零开始的pytorch小白使用指北

Pytorch是什么

PyTorch是一个开源的Python机器学习库，提供两个高级功能：

1. 具有强大的GPU加速的张量计算（如NumPy）

Pytorch向numpy做了广泛兼容，很多numpy的函数接口在pytorch下同样适用，
因此在安装pytorch和numpy时要注意版本兼容。

2. 包含自动求导系统的深度神经网络

Pytorch搭建神经网络

>>> import torch  
>>> import numpy 

数据准备

tensor：张量，用于做矩阵的运算，十分适合在GPU上跑。
构建张量

>>> a = torch.tensor([[1,2,3],[4,5,6]],dtype=float)	# 创建一个指定元素的张量
>>> a.size()	
torch.Size([2, 3])
>>> a	
tensor([[1., 2., 3.],
        [4., 5., 6.]], dtype=torch.float64)

>>> b = torch.full((2,3),3.14)	# 创建一个全为某一数值的张量，可以作为初始化矩阵
>>> b
tensor([[3.1400, 3.1400, 3.1400],
        [3.1400, 3.1400, 3.1400]])

>>> c = torch.empty(2,3) # 返回一个未初始化的指定形状的张量，每次的运行结果都不同
>>> c
tensor([[3.6893e+19, 2.1425e+00, 5.9824e-38],
        [1.7449e-39, 6.1124e-38, 1.4013e-45]])

张量运算

>>> #加法运算
>>> a+b
tensor([[4.1400, 5.1400, 6.1400],
        [7.1400, 8.1400, 9.1400]])
        
>>> torch.add(a,b,out=c) # 将指定张量作为输出，若不指定直接返回该值
>>> c
tensor([[4.1400, 5.1400, 6.1400],
        [7.1400, 8.1400, 9.1400]])
        
>>> a.add(b) #返回一个新的张量，原张量不变
>>> a
tensor([[1., 2., 3.],
        [4., 5., 6.]], dtype=torch.float64)
        
>>> a.add_(b) # 张量操作加'_'后，原张量值改变
>>> a
tensor([[4.1400, 5.1400, 6.1400],
        [7.1400, 8.1400, 9.1400]], dtype=torch.float64)

# 乘法运算
>>> x = torch.rand(3,5)
>>> y = torch.rand(3,5)
>>> z = torch.rand(5,3)
>>> torch.mul(x,y) # 矩阵点乘
tensor([[0.5100, 0.2523, 0.3176, 0.7665, 0.7989],
        [0.3799, 0.0779, 0.1352, 0.5991, 0.0338],
        [0.0375, 0.1444, 0.6965, 0.0810, 0.1836]])
>>> torch.matmul(x,z) #矩阵叉乘
tensor([[0.9843, 2.0296, 1.3304],
        [0.7903, 1.4374, 1.1705],
        [1.0530, 1.9607, 0.9520]])

# 常用运算
>>> a = torch.tensor([[1,2,3],[4,5,6]])
>>> torch.max(a) # 所有张量元素最大值
tensor(6)
>>> torch.max(a,0) # 沿着第0维求最大值
torch.return_types.max(values=tensor([4, 5, 6]),indices=tensor([1, 1, 1]))
>>> torch.max(a,1) # 沿着第1维求最大值
torch.return_types.max(values=tensor([3, 6]),indices=tensor([2, 2]))
>>> # min,mean,sum,prod同理

numpy桥
numpy.array和torch.tensor可以互相转化

>>> a = np.ones(3)
>>> a
array([1., 1., 1.])
>>> aa = torch.from_numpy(a) # array转为tensor
>>> aa
tensor([1., 1., 1.], dtype=torch.float64)
>>> b = torch.ones(3)
>>> b
tensor([1., 1., 1.])
>>> bb = b.numpy() # tensor转为array
>>> bb
array([1., 1., 1.], dtype=float32)

构建data_iterator

在这里插入代码片

模型准备

Module：一个神经网络的层次，可以直接调用全连接层，卷积层，等等神经网络。被自定义的神经网络类所继承。
自定义一个神经网络class通常包括两部分：__init__定义网络层，forward函数定义网络计算函数。

# Example 1
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool  = nn.MaxPool2d(2,2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1   = nn.Linear(16*5*5, 120)
        self.fc2   = nn.Linear(120, 84)
        self.fc3   = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16*5*5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

NetModel = Net()

# Example 2
class DynamicNet(torch.nn.Module):
    def __init__(self, D_in, H, D_out):
        super(DynamicNet, self).__init__()
        self.input_linear = torch.nn.Linear(D_in, H)
        self.middle_linear = torch.nn.Linear(H, H)
        self.output_linear = torch.nn.Linear(H, D_out)

    def forward(self, x):
        h_relu = self.input_linear(x).clamp(min=0)
        for _ in range(random.randint(0, 3)):
            h_relu = self.middle_linear(h_relu).clamp(min=0)
        y_pred = self.output_linear(h_relu)
        return y_pred
        
DynamicModel = DynamicNet(D_in, H, D_out)

模型训练

模型训练的主题代码由2层循环、2个定义和5句函数组成：

# 两个定义：
optimizer = 
criterion = 
# 两层循环:
for epoch in num_epochs:
	for input in batch_iters:
		# 5句函数
		outputs = model(input)
        model.zero_grad()
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

# Example 1
# 定义代价函数和优化器
criterion = nn.CrossEntropyLoss() # use a Classification Cross-Entropy loss
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
# 训练 epoch × batches_per_epoch 轮
for epoch in range(2): 
    running_loss = 0.0
    # 遍历trainloader时每次返回一个batch
    for i, data in enumerate(trainloader, 0):
        inputs, labels = data
        # wrap them in Variable
        inputs, labels = Variable(inputs), Variable(labels)
   
        # zero the parameter gradients
        optimizer.zero_grad()
        
        # forward + backward + optimize
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()        
        optimizer.step()
        running_loss += loss.data[0]
        
        # print statistics
        if i % 2000 == 1999: 
            print('[%d, %5d] loss: %.3f' % (epoch+1, i+1, running_loss / 2000))
            running_loss = 0.0

# Example 2 简化版
criterion = torch.nn.MSELoss(reduction='sum')
optimizer = torch.optim.SGD(model.parameters(), lr=1e-4, momentum=0.9)

for t in range(500):
    y_pred = model(x)
    loss = criterion(y_pred, y)
    print(t, loss.item())
    # 清零梯度，反向传播，更新权重 
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

# Example3
 optimizer = torch.optim.Adam(model.parameters(), lr=config.learning_rate)
 model.train()
 for epoch in range(config.num_epochs):
     print('Epoch [{}/{}]'.format(epoch + 1, config.num_epochs))
     for i, (trains, labels) in enumerate(train_iter):
         outputs = model(trains)
         model.zero_grad()
         loss = F.cross_entropy(outputs, labels.squeeze(1))
         loss.backward()
         optimizer.step()

模型评价

一个循环+2句计算+n个评价函数：

# 一个循环：test_iters
for texts, labels in test_iter:
    # 2句计算：outputs = model(inputs)
    outputs = net(Variable(images))
    _, predicts = torch.max(outputs.data, 1)
# n个评价函数：用labels和predicts计算metrcs库的各种评价指标

# Example1
correct = 0
total = 0
# 一个循环：test_iters
for data in testloader:
    images, labels = data
    # 2句计算：outputs = model(inputs)
    outputs = net(Variable(images))
    _, predicted = torch.max(outputs.data, 1)
    total += labels.size(0)
    correct += (predicted == labels).sum()
print('Accuracy of the network on the 10000 test images: %d %%' % (100 * correct / total))

# Example 2
model.load_state_dict(torch.load(config.save_path))
model.eval()
loss_total = 0
predict_all = np.array([], dtype=int)
labels_all = np.array([], dtype=int)
with torch.no_grad():# 不自动求导，节约内存
    for texts, labels in data_iter:
        outputs = model(texts)
        predic = torch.max(outputs.data, 1)[1].cpu().numpy()
        loss = F.cross_entropy(outputs, labels.squeeze(1))
        loss_total += loss
        labels = labels.data.cpu().numpy()
        labels_all = np.append(labels_all, labels)
        predict_all = np.append(predict_all, predic)

acc = metrics.accuracy_score(labels_all, predict_all)
report = metrics.classification_report(labels_all, predict_all, 
confusion = metrics.confusion_matrix(labels_all, predict_all)
print (acc, loss_total / len(data_iter), report, confusion)

使用GPU训练神经网络

# 使用GPU训练时，需要将model和tensor转至指定device
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = BertClassifier(config.bert_path)
model = model.to(device)
tensor = torch.tensor(x_train).to(torch.float32)
tensor = tensor.to(device)

Pytorch和TensorFlow的区别

1. Pytorch构建动态计算图，TensorFlow构建静态计算图，因此pytorch更为灵活一些；
   

计算图示例

2. TensorFlow在GPU上的计算性能更高一些，在CPU上二者类似。