优化模型参数#

现在已经拥有了模型和数据,是时候通过在数据上优化其参数来训练、验证和测试模型了。训练模型是一个迭代过程;在每次迭代中,模型会对输出做出猜测,计算其猜测中的误差(损失),收集误差相对于其参数的导数(正如我们在先前部分看到的),并使用梯度下降法 优化 这些参数。想要更详细地了解这个过程,请查看这个关于反向传播的视频,由3Blue1Brown提供。

先决条件代码#

从之前的部分加载代码,包括数据集与DataLoaders构建模型

from set_env import temp_dir
项目根目录:/media/pc/data/lxw/ai/torch-book
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor

training_data = datasets.FashionMNIST(
    root=temp_dir/"data",
    train=True,
    download=True,
    transform=ToTensor()
)

test_data = datasets.FashionMNIST(
    root=temp_dir/"data",
    train=False,
    download=True,
    transform=ToTensor()
)

train_dataloader = DataLoader(training_data, batch_size=64)
test_dataloader = DataLoader(test_data, batch_size=64)

class NeuralNetwork(nn.Module):
    def __init__(self):
        super().__init__()
        self.flatten = nn.Flatten()
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(28*28, 512),
            nn.ReLU(),
            nn.Linear(512, 512),
            nn.ReLU(),
            nn.Linear(512, 10),
        )

    def forward(self, x):
        x = self.flatten(x)
        logits = self.linear_relu_stack(x)
        return logits

model = NeuralNetwork()

超参数#

超参数是可调参数,允许控制模型优化过程。不同的超参数值会影响模型训练和收敛速度(了解更多 关于超参数调优)

为训练定义了以下超参数:

  • Epoch 数量 - 遍历数据集的次数

  • 批量大小 - 在更新参数之前通过网络传播的数据样本数量

  • 学习率 - 在每个批次/epoch 更新模型参数的幅度。较小的值会导致学习速度缓慢,而较大的值可能会导致训练过程中出现不可预测的行为。

learning_rate = 1e-3
batch_size = 64
epochs = 5

循环优化#

一旦我们设置了超参数,就可以通过循环优化来训练和优化我们的模型。优化循环的每次迭代称为一个 epoch

每个 epoch 包含两个主要部分:

  • 训练循环:遍历训练数据集,尝试收敛到最优参数。

  • 验证/测试循环:遍历测试数据集,检查模型性能是否有所提高。

简要熟悉一下训练循环中使用的一些概念。跳到 完整实现 查看循环优化的完整实现。

损失函数#

当给定一些训练数据时,未训练的网络可能不会给出正确的答案。损失函数 衡量获得结果与目标值之间的差异程度,希望在训练过程中最小化损失函数。为了计算损失,使用给定数据样本的输入进行预测,并将其与真实数据标签值进行比较。

常见的损失函数包括 torch.nn.MSELoss(均方误差)用于回归任务,以及 torch.nn.NLLLoss(负对数似然)用于分类任务。torch.nn.CrossEntropyLoss 结合了 torch.nn.LogSoftmaxtorch.nn.NLLLoss

将模型的输出 logits 传递给 torch.nn.CrossEntropyLoss,它将对 logits 进行归一化并计算预测误差。

# Initialize the loss function
loss_fn = nn.CrossEntropyLoss()

优化器#

优化是调整模型参数以在每个训练步骤中减少模型误差的过程。优化算法 定义了如何执行此过程(在本例中,使用随机梯度下降)。所有优化逻辑都封装在 optimizer 对象中。在这里,使用 SGD 优化器;此外,PyTorch 中还有许多 不同的优化器,例如 ADAM 和 RMSProp,它们适用于不同类型的模型和数据。

通过注册需要训练的模型参数并传入学习率超参数来初始化优化器。

optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)

在训练循环内部,优化分为三个步骤:

  • 调用 optimizer.zero_grad() 来重置模型参数的梯度。默认情况下,梯度会累加;为了避免重复计算,我们在每次迭代时显式地将它们归零。

  • 通过调用 loss.backward() 反向传播预测损失。PyTorch 将损失相对于每个参数的梯度存储起来。

  • 一旦我们有了梯度,我们调用 optimizer.step() 来根据反向传播中收集的梯度调整参数。

完整实现#

定义 train_loop 来循环执行优化代码,并定义 test_loop 来评估模型在测试数据上的性能。

def train_loop(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)
    # Set the model to training mode - important for batch normalization and dropout layers
    # Unnecessary in this situation but added for best practices
    model.train()
    for batch, (X, y) in enumerate(dataloader):
        # Compute prediction and loss
        pred = model(X)
        loss = loss_fn(pred, y)

        # Backpropagation
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

        if batch % 100 == 0:
            loss, current = loss.item(), batch * batch_size + len(X)
            print(f"loss: {loss:>7f}  [{current:>5d}/{size:>5d}]")


def test_loop(dataloader, model, loss_fn):
    # Set the model to evaluation mode - important for batch normalization and dropout layers
    # Unnecessary in this situation but added for best practices
    model.eval()
    size = len(dataloader.dataset)
    num_batches = len(dataloader)
    test_loss, correct = 0, 0

    # Evaluating the model with torch.no_grad() ensures that no gradients are computed during test mode
    # also serves to reduce unnecessary gradient computations and memory usage for tensors with requires_grad=True
    with torch.no_grad():
        for X, y in dataloader:
            pred = model(X)
            test_loss += loss_fn(pred, y).item()
            correct += (pred.argmax(1) == y).type(torch.float).sum().item()

    test_loss /= num_batches
    correct /= size
    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")

初始化损失函数和优化器,并将其传递给 train_looptest_loop。可以随意增加 epoch 的数量以跟踪模型性能的提升。

loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)

epochs = 10
for t in range(epochs):
    print(f"Epoch {t+1}\n-------------------------------")
    train_loop(train_dataloader, model, loss_fn, optimizer)
    test_loop(test_dataloader, model, loss_fn)
print("Done!")
Hide code cell output
Epoch 1
-------------------------------
loss: 2.299955  [   64/60000]
loss: 2.289047  [ 6464/60000]
loss: 2.262702  [12864/60000]
loss: 2.262147  [19264/60000]
loss: 2.255074  [25664/60000]
loss: 2.209852  [32064/60000]
loss: 2.228119  [38464/60000]
loss: 2.189891  [44864/60000]
loss: 2.186696  [51264/60000]
loss: 2.157851  [57664/60000]
Test Error: 
 Accuracy: 43.2%, Avg loss: 2.150610 

Epoch 2
-------------------------------
loss: 2.160350  [   64/60000]
loss: 2.152020  [ 6464/60000]
loss: 2.087391  [12864/60000]
loss: 2.109184  [19264/60000]
loss: 2.062347  [25664/60000]
loss: 1.993510  [32064/60000]
loss: 2.032360  [38464/60000]
loss: 1.952619  [44864/60000]
loss: 1.954293  [51264/60000]
loss: 1.886026  [57664/60000]
Test Error: 
 Accuracy: 60.6%, Avg loss: 1.879761 

Epoch 3
-------------------------------
loss: 1.912941  [   64/60000]
loss: 1.882683  [ 6464/60000]
loss: 1.757799  [12864/60000]
loss: 1.807219  [19264/60000]
loss: 1.694405  [25664/60000]
loss: 1.644480  [32064/60000]
loss: 1.674132  [38464/60000]
loss: 1.576804  [44864/60000]
loss: 1.595409  [51264/60000]
loss: 1.493224  [57664/60000]
Test Error: 
 Accuracy: 62.9%, Avg loss: 1.508747 

Epoch 4
-------------------------------
loss: 1.576287  [   64/60000]
loss: 1.541384  [ 6464/60000]
loss: 1.386807  [12864/60000]
loss: 1.465601  [19264/60000]
loss: 1.344606  [25664/60000]
loss: 1.341592  [32064/60000]
loss: 1.360759  [38464/60000]
loss: 1.288418  [44864/60000]
loss: 1.315813  [51264/60000]
loss: 1.218872  [57664/60000]
Test Error: 
 Accuracy: 64.0%, Avg loss: 1.244145 

Epoch 5
-------------------------------
loss: 1.322750  [   64/60000]
loss: 1.304346  [ 6464/60000]
loss: 1.136508  [12864/60000]
loss: 1.245479  [19264/60000]
loss: 1.122106  [25664/60000]
loss: 1.148881  [32064/60000]
loss: 1.172815  [38464/60000]
loss: 1.111189  [44864/60000]
loss: 1.142142  [51264/60000]
loss: 1.061584  [57664/60000]
Test Error: 
 Accuracy: 65.1%, Avg loss: 1.081436 

Epoch 6
-------------------------------
loss: 1.154250  [   64/60000]
loss: 1.156330  [ 6464/60000]
loss: 0.972618  [12864/60000]
loss: 1.107574  [19264/60000]
loss: 0.986492  [25664/60000]
loss: 1.020281  [32064/60000]
loss: 1.059144  [38464/60000]
loss: 0.999622  [44864/60000]
loss: 1.029244  [51264/60000]
loss: 0.964120  [57664/60000]
Test Error: 
 Accuracy: 66.0%, Avg loss: 0.977057 

Epoch 7
-------------------------------
loss: 1.037883  [   64/60000]
loss: 1.060785  [ 6464/60000]
loss: 0.861099  [12864/60000]
loss: 1.015254  [19264/60000]
loss: 0.901760  [25664/60000]
loss: 0.929478  [32064/60000]
loss: 0.985907  [38464/60000]
loss: 0.928371  [44864/60000]
loss: 0.951404  [51264/60000]
loss: 0.899121  [57664/60000]
Test Error: 
 Accuracy: 66.9%, Avg loss: 0.906147 

Epoch 8
-------------------------------
loss: 0.952631  [   64/60000]
loss: 0.994779  [ 6464/60000]
loss: 0.781947  [12864/60000]
loss: 0.949366  [19264/60000]
loss: 0.845234  [25664/60000]
loss: 0.862846  [32064/60000]
loss: 0.935118  [38464/60000]
loss: 0.881270  [44864/60000]
loss: 0.894825  [51264/60000]
loss: 0.852345  [57664/60000]
Test Error: 
 Accuracy: 68.1%, Avg loss: 0.855028 

Epoch 9
-------------------------------
loss: 0.887047  [   64/60000]
loss: 0.945449  [ 6464/60000]
loss: 0.722958  [12864/60000]
loss: 0.899695  [19264/60000]
loss: 0.804949  [25664/60000]
loss: 0.812118  [32064/60000]
loss: 0.896828  [38464/60000]
loss: 0.848763  [44864/60000]
loss: 0.851984  [51264/60000]
loss: 0.816333  [57664/60000]
Test Error: 
 Accuracy: 69.3%, Avg loss: 0.816232 

Epoch 10
-------------------------------
loss: 0.834691  [   64/60000]
loss: 0.905861  [ 6464/60000]
loss: 0.677137  [12864/60000]
loss: 0.861224  [19264/60000]
loss: 0.774445  [25664/60000]
loss: 0.772846  [32064/60000]
loss: 0.866035  [38464/60000]
loss: 0.825091  [44864/60000]
loss: 0.818532  [51264/60000]
loss: 0.787082  [57664/60000]
Test Error: 
 Accuracy: 70.5%, Avg loss: 0.785396 

Done!

进一步阅读#