Shortcuts

知识蒸馏教程

创建于:2023年8月22日 | 最后更新:2024年7月30日 | 最后验证:2024年11月5日

作者: Alexandros Chariton

知识蒸馏是一种技术,它使得知识能够从大型、计算成本高的模型转移到较小的模型,而不会失去有效性。这使得在性能较低的硬件上部署成为可能,从而使评估更快、更高效。

在本教程中,我们将运行一系列实验,旨在通过使用更强大的网络作为教师来提高轻量级神经网络的准确性。 轻量级网络的计算成本和速度将保持不变,我们的干预仅关注其权重,而不影响其前向传播。 该技术的应用可以在无人机或手机等设备中找到。 在本教程中,我们不使用任何外部包,因为我们所需的一切都在torchtorchvision中可用。

在本教程中,您将学习:

  • 如何修改模型类以提取隐藏表示并将其用于进一步计算

  • 如何在PyTorch中修改常规的训练循环,以在分类的交叉熵之上包含额外的损失

  • 如何通过使用更复杂的模型作为教师来提高轻量级模型的性能

先决条件

  • 1个GPU,4GB内存

  • PyTorch v2.0 或更高版本

  • CIFAR-10 数据集(由脚本下载并保存在名为 /data 的目录中)

import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
import torchvision.datasets as datasets

# Check if GPU is available, and if not, use the CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

加载CIFAR-10

CIFAR-10 是一个包含十个类别的流行图像数据集。我们的目标是为每个输入图像预测以下类别之一。

../_static/img/cifar10.png

CIFAR-10 图像示例

输入的图像是RGB格式的,因此它们有3个通道,并且是32x32像素的。基本上,每张图像由3 x 32 x 32 = 3072个数字描述,范围从0到255。 在神经网络中,一个常见的做法是对输入进行归一化,这样做有多种原因, 包括避免常用激活函数中的饱和以及增加数值稳定性。 我们的归一化过程包括减去均值并除以每个通道的标准差。 张量“mean=[0.485, 0.456, 0.406]”和“std=[0.229, 0.224, 0.225]”已经被计算出来, 它们代表了预定义的CIFAR-10训练集子集中每个通道的均值和标准差。 注意我们如何将这些值也用于测试集,而不从头重新计算均值和标准差。 这是因为网络是在通过减去和除以上述数字产生的特征上进行训练的,我们希望保持一致性。 此外,在现实生活中,我们将无法计算测试集的均值和标准差,因为根据我们的假设,这些数据在那个时候是不可访问的。

作为总结,我们通常将这个保留集称为验证集,并在优化模型在验证集上的性能后使用一个单独的集合,称为测试集。这样做是为了避免基于单一指标的贪婪和有偏优化来选择模型。

# Below we are preprocessing data for CIFAR-10. We use an arbitrary batch size of 128.
transforms_cifar = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

# Loading the CIFAR-10 dataset:
train_dataset = datasets.CIFAR10(root='./data', train=True, download=True, transform=transforms_cifar)
test_dataset = datasets.CIFAR10(root='./data', train=False, download=True, transform=transforms_cifar)
Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to ./data/cifar-10-python.tar.gz

  0%|          | 0.00/170M [00:00<?, ?B/s]
  0%|          | 492k/170M [00:00<00:34, 4.91MB/s]
  5%|4         | 8.16M/170M [00:00<00:03, 47.1MB/s]
 12%|#1        | 19.9M/170M [00:00<00:01, 79.1MB/s]
 19%|#8        | 31.7M/170M [00:00<00:01, 94.2MB/s]
 25%|##5       | 43.4M/170M [00:00<00:01, 103MB/s]
 32%|###2      | 55.2M/170M [00:00<00:01, 108MB/s]
 39%|###9      | 66.9M/170M [00:00<00:00, 111MB/s]
 46%|####6     | 78.6M/170M [00:00<00:00, 113MB/s]
 53%|#####3    | 90.4M/170M [00:00<00:00, 114MB/s]
 60%|#####9    | 102M/170M [00:01<00:00, 115MB/s]
 67%|######6   | 114M/170M [00:01<00:00, 116MB/s]
 74%|#######3  | 126M/170M [00:01<00:00, 116MB/s]
 81%|########  | 137M/170M [00:01<00:00, 117MB/s]
 88%|########7 | 149M/170M [00:01<00:00, 117MB/s]
 94%|#########4| 161M/170M [00:01<00:00, 117MB/s]
100%|##########| 170M/170M [00:01<00:00, 108MB/s]
Extracting ./data/cifar-10-python.tar.gz to ./data
Files already downloaded and verified

注意

本节仅适用于对快速结果感兴趣的CPU用户。仅当您对小型实验感兴趣时使用此选项。请记住,使用任何GPU代码都应该运行得相当快。仅从训练/测试数据集中选择前num_images_to_keep张图像

#from torch.utils.data import Subset
#num_images_to_keep = 2000
#train_dataset = Subset(train_dataset, range(min(num_images_to_keep, 50_000)))
#test_dataset = Subset(test_dataset, range(min(num_images_to_keep, 10_000)))
#Dataloaders
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=128, shuffle=True, num_workers=2)
test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=128, shuffle=False, num_workers=2)

定义模型类和实用函数

接下来,我们需要定义我们的模型类。这里需要设置几个用户定义的参数。我们使用两种不同的架构,保持实验中过滤器的数量固定,以确保公平比较。 两种架构都是卷积神经网络(CNNs),具有不同数量的卷积层作为特征提取器,后面跟着一个有10个类别的分类器。 对于学生来说,过滤器和神经元的数量较少。

# Deeper neural network class to be used as teacher:
class DeepNN(nn.Module):
    def __init__(self, num_classes=10):
        super(DeepNN, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 128, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(128, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(64, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(64, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )
        self.classifier = nn.Sequential(
            nn.Linear(2048, 512),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(512, num_classes)
        )

    def forward(self, x):
        x = self.features(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x

# Lightweight neural network class to be used as student:
class LightNN(nn.Module):
    def __init__(self, num_classes=10):
        super(LightNN, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(16, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )
        self.classifier = nn.Sequential(
            nn.Linear(1024, 256),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(256, num_classes)
        )

    def forward(self, x):
        x = self.features(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x

我们使用2个函数来帮助我们生成和评估原始分类任务的结果。 一个函数称为train,并接受以下参数:

  • model: 通过此函数训练的模型实例(更新其权重)。

  • train_loader: 我们在上面定义了train_loader,它的工作是将数据输入到模型中。

  • epochs: 我们循环数据集的次数。

  • learning_rate: 学习率决定了我们向收敛迈进的步长应该有多大。步长太大或太小都可能有害。

  • device: 确定运行工作负载的设备。根据可用性,可以是CPU或GPU。

我们的测试函数类似,但它将使用test_loader来加载测试集中的图像。

../_static/img/knowledge_distillation/ce_only.png

使用交叉熵训练两个网络。学生网络将作为基线:

def train(model, train_loader, epochs, learning_rate, device):
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=learning_rate)

    model.train()

    for epoch in range(epochs):
        running_loss = 0.0
        for inputs, labels in train_loader:
            # inputs: A collection of batch_size images
            # labels: A vector of dimensionality batch_size with integers denoting class of each image
            inputs, labels = inputs.to(device), labels.to(device)

            optimizer.zero_grad()
            outputs = model(inputs)

            # outputs: Output of the network for the collection of images. A tensor of dimensionality batch_size x num_classes
            # labels: The actual labels of the images. Vector of dimensionality batch_size
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()

            running_loss += loss.item()

        print(f"Epoch {epoch+1}/{epochs}, Loss: {running_loss / len(train_loader)}")

def test(model, test_loader, device):
    model.to(device)
    model.eval()

    correct = 0
    total = 0

    with torch.no_grad():
        for inputs, labels in test_loader:
            inputs, labels = inputs.to(device), labels.to(device)

            outputs = model(inputs)
            _, predicted = torch.max(outputs.data, 1)

            total += labels.size(0)
            correct += (predicted == labels).sum().item()

    accuracy = 100 * correct / total
    print(f"Test Accuracy: {accuracy:.2f}%")
    return accuracy

交叉熵运行

为了可重复性,我们需要设置torch的手动种子。我们使用不同的方法训练网络,因此为了公平比较,用相同的权重初始化网络是有意义的。 首先使用交叉熵训练教师网络:

torch.manual_seed(42)
nn_deep = DeepNN(num_classes=10).to(device)
train(nn_deep, train_loader, epochs=10, learning_rate=0.001, device=device)
test_accuracy_deep = test(nn_deep, test_loader, device)

# Instantiate the lightweight network:
torch.manual_seed(42)
nn_light = LightNN(num_classes=10).to(device)
Epoch 1/10, Loss: 1.3366997722164748
Epoch 2/10, Loss: 0.8720864758772009
Epoch 3/10, Loss: 0.6820237918583023
Epoch 4/10, Loss: 0.5375950956893394
Epoch 5/10, Loss: 0.41490358377204223
Epoch 6/10, Loss: 0.3123219789141584
Epoch 7/10, Loss: 0.22101545549185989
Epoch 8/10, Loss: 0.17098309014878615
Epoch 9/10, Loss: 0.13455525941460791
Epoch 10/10, Loss: 0.12078842208208636
Test Accuracy: 75.45%

我们实例化了一个更轻量级的网络模型来比较它们的性能。 反向传播对权重初始化很敏感, 所以我们需要确保这两个网络具有完全相同的初始化。

torch.manual_seed(42)
new_nn_light = LightNN(num_classes=10).to(device)

为了确保我们已经创建了第一个网络的副本,我们检查其第一层的范数。 如果匹配,那么我们可以安全地得出结论,网络确实是相同的。

# Print the norm of the first layer of the initial lightweight model
print("Norm of 1st layer of nn_light:", torch.norm(nn_light.features[0].weight).item())
# Print the norm of the first layer of the new lightweight model
print("Norm of 1st layer of new_nn_light:", torch.norm(new_nn_light.features[0].weight).item())
Norm of 1st layer of nn_light: 2.327361822128296
Norm of 1st layer of new_nn_light: 2.327361822128296

打印每个模型中的参数总数:

total_params_deep = "{:,}".format(sum(p.numel() for p in nn_deep.parameters()))
print(f"DeepNN parameters: {total_params_deep}")
total_params_light = "{:,}".format(sum(p.numel() for p in nn_light.parameters()))
print(f"LightNN parameters: {total_params_light}")
DeepNN parameters: 1,186,986
LightNN parameters: 267,738

使用交叉熵损失训练和测试轻量级网络:

train(nn_light, train_loader, epochs=10, learning_rate=0.001, device=device)
test_accuracy_light_ce = test(nn_light, test_loader, device)
Epoch 1/10, Loss: 1.466049101346594
Epoch 2/10, Loss: 1.1519653670623173
Epoch 3/10, Loss: 1.0232561651398153
Epoch 4/10, Loss: 0.9235453337354733
Epoch 5/10, Loss: 0.8479179534156
Epoch 6/10, Loss: 0.7824301378196462
Epoch 7/10, Loss: 0.7184310383199121
Epoch 8/10, Loss: 0.6588469929707325
Epoch 9/10, Loss: 0.6075488568266945
Epoch 10/10, Loss: 0.556159371533967
Test Accuracy: 70.15%

正如我们所看到的,基于测试准确率,我们现在可以将用作教师的深层网络与假设为学生网络的轻量级网络进行比较。到目前为止,我们的学生还没有干预教师,因此这个表现是由学生自己实现的。 到目前为止的指标可以通过以下行看到:

print(f"Teacher accuracy: {test_accuracy_deep:.2f}%")
print(f"Student accuracy: {test_accuracy_light_ce:.2f}%")
Teacher accuracy: 75.45%
Student accuracy: 70.15%

知识蒸馏运行

现在让我们尝试通过结合教师网络来提高学生网络的测试准确率。 知识蒸馏是一种直接的技术来实现这一点, 基于两个网络都输出类别的概率分布这一事实。 因此,这两个网络共享相同数量的输出神经元。 该方法通过在传统的交叉熵损失中加入额外的损失来工作, 这是基于教师网络的softmax输出。 假设是,一个经过适当训练的教师网络的输出激活携带了额外的信息,这些信息可以在训练过程中被学生网络利用。 原始工作表明,利用软目标中较小概率的比率可以帮助实现深度神经网络的基本目标, 即在数据上创建相似性结构,其中相似的对象被映射得更近。 例如,在CIFAR-10中,如果卡车的轮子存在,它可能会被误认为是汽车或飞机, 但不太可能被误认为是狗。 因此,假设有价值的信息不仅存在于经过适当训练的模型的顶部预测中,而且存在于整个输出分布中是有意义的。 然而,仅靠交叉熵并不能充分利用这些信息,因为非预测类别的激活 往往非常小,以至于传播的梯度不会有意义地改变权重来构建这个理想的向量空间。

在我们继续定义第一个引入师生动态的辅助函数时,我们需要包含一些额外的参数:

  • T: 温度控制输出分布的平滑度。较大的T会导致分布更平滑,因此较小的概率会得到更大的提升。

  • soft_target_loss_weight: 分配给我们将要包含的额外目标的权重。

  • ce_loss_weight: 分配给交叉熵的权重。调整这些权重可以推动网络优化任一目标。

../_static/img/knowledge_distillation/distillation_output_loss.png

蒸馏损失是从网络的logits计算的。它只返回梯度给学生:

def train_knowledge_distillation(teacher, student, train_loader, epochs, learning_rate, T, soft_target_loss_weight, ce_loss_weight, device):
    ce_loss = nn.CrossEntropyLoss()
    optimizer = optim.Adam(student.parameters(), lr=learning_rate)

    teacher.eval()  # Teacher set to evaluation mode
    student.train() # Student to train mode

    for epoch in range(epochs):
        running_loss = 0.0
        for inputs, labels in train_loader:
            inputs, labels = inputs.to(device), labels.to(device)

            optimizer.zero_grad()

            # Forward pass with the teacher model - do not save gradients here as we do not change the teacher's weights
            with torch.no_grad():
                teacher_logits = teacher(inputs)

            # Forward pass with the student model
            student_logits = student(inputs)

            #Soften the student logits by applying softmax first and log() second
            soft_targets = nn.functional.softmax(teacher_logits / T, dim=-1)
            soft_prob = nn.functional.log_softmax(student_logits / T, dim=-1)

            # Calculate the soft targets loss. Scaled by T**2 as suggested by the authors of the paper "Distilling the knowledge in a neural network"
            soft_targets_loss = torch.sum(soft_targets * (soft_targets.log() - soft_prob)) / soft_prob.size()[0] * (T**2)

            # Calculate the true label loss
            label_loss = ce_loss(student_logits, labels)

            # Weighted sum of the two losses
            loss = soft_target_loss_weight * soft_targets_loss + ce_loss_weight * label_loss

            loss.backward()
            optimizer.step()

            running_loss += loss.item()

        print(f"Epoch {epoch+1}/{epochs}, Loss: {running_loss / len(train_loader)}")

# Apply ``train_knowledge_distillation`` with a temperature of 2. Arbitrarily set the weights to 0.75 for CE and 0.25 for distillation loss.
train_knowledge_distillation(teacher=nn_deep, student=new_nn_light, train_loader=train_loader, epochs=10, learning_rate=0.001, T=2, soft_target_loss_weight=0.25, ce_loss_weight=0.75, device=device)
test_accuracy_light_ce_and_kd = test(new_nn_light, test_loader, device)

# Compare the student test accuracy with and without the teacher, after distillation
print(f"Teacher accuracy: {test_accuracy_deep:.2f}%")
print(f"Student accuracy without teacher: {test_accuracy_light_ce:.2f}%")
print(f"Student accuracy with CE + KD: {test_accuracy_light_ce_and_kd:.2f}%")
Epoch 1/10, Loss: 2.3962801237545355
Epoch 2/10, Loss: 1.87831118161721
Epoch 3/10, Loss: 1.6540942881113427
Epoch 4/10, Loss: 1.4959803764777415
Epoch 5/10, Loss: 1.367971143454237
Epoch 6/10, Loss: 1.2519247448048019
Epoch 7/10, Loss: 1.1570622474336258
Epoch 8/10, Loss: 1.0719402747995712
Epoch 9/10, Loss: 0.9970421949615869
Epoch 10/10, Loss: 0.9293939061177051
Test Accuracy: 70.75%
Teacher accuracy: 75.45%
Student accuracy without teacher: 70.15%
Student accuracy with CE + KD: 70.75%

余弦损失最小化运行

请随意调整控制softmax函数柔和度的温度参数和损失系数。 在神经网络中,很容易将额外的损失函数包含到主要目标中,以实现更好的泛化等目标。 让我们尝试为学生包含一个目标,但现在让我们关注他们的隐藏状态而不是输出层。 我们的目标是通过包含一个简单的损失函数,将信息从教师的表示传递给学生, 其最小化意味着随后传递给分类器的扁平向量随着损失的减少而变得更加相似。 当然,教师不会更新其权重,因此最小化仅取决于学生的权重。 这种方法的基本原理是,我们假设教师模型具有更好的内部表示,而学生没有外部干预是不可能达到的,因此我们人为地推动学生模仿教师的内部表示。 然而,这是否最终会帮助学生并不直接,因为推动轻量级网络达到这一点可能是好事,假设我们找到了一个导致更好测试准确性的内部表示, 但它也可能是有害的,因为网络具有不同的架构,学生没有与教师相同的学习能力。 换句话说,这两个向量,学生和教师的向量,没有理由在每个组件上匹配。 学生可以达到一个教师表示的排列,这将同样有效。 尽管如此,我们仍然可以运行一个快速实验来找出这种方法的影响。 我们将使用CosineEmbeddingLoss,其公式如下:

../_static/img/knowledge_distillation/cosine_embedding_loss.png

余弦嵌入损失的公式

显然,我们首先需要解决一个问题。 当我们将蒸馏应用于输出层时,我们提到两个网络具有相同数量的神经元,等于类别数量。 然而,对于卷积层之后的层来说,情况并非如此。在这里,教师在最终卷积层展平后比学生拥有更多的神经元。 我们的损失函数接受两个维度相等的向量作为输入,因此我们需要以某种方式匹配它们。我们将通过在教师的卷积层之后包含一个平均池化层来解决这个问题,以将其维度减少到与学生相匹配。

为了继续,我们将修改我们的模型类,或创建新的模型类。 现在,forward函数不仅返回网络的logits,还返回卷积层后的扁平化隐藏表示。我们为修改后的教师模型包含了上述的池化操作。

class ModifiedDeepNNCosine(nn.Module):
    def __init__(self, num_classes=10):
        super(ModifiedDeepNNCosine, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 128, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(128, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(64, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(64, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )
        self.classifier = nn.Sequential(
            nn.Linear(2048, 512),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(512, num_classes)
        )

    def forward(self, x):
        x = self.features(x)
        flattened_conv_output = torch.flatten(x, 1)
        x = self.classifier(flattened_conv_output)
        flattened_conv_output_after_pooling = torch.nn.functional.avg_pool1d(flattened_conv_output, 2)
        return x, flattened_conv_output_after_pooling

# Create a similar student class where we return a tuple. We do not apply pooling after flattening.
class ModifiedLightNNCosine(nn.Module):
    def __init__(self, num_classes=10):
        super(ModifiedLightNNCosine, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(16, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )
        self.classifier = nn.Sequential(
            nn.Linear(1024, 256),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(256, num_classes)
        )

    def forward(self, x):
        x = self.features(x)
        flattened_conv_output = torch.flatten(x, 1)
        x = self.classifier(flattened_conv_output)
        return x, flattened_conv_output

# We do not have to train the modified deep network from scratch of course, we just load its weights from the trained instance
modified_nn_deep = ModifiedDeepNNCosine(num_classes=10).to(device)
modified_nn_deep.load_state_dict(nn_deep.state_dict())

# Once again ensure the norm of the first layer is the same for both networks
print("Norm of 1st layer for deep_nn:", torch.norm(nn_deep.features[0].weight).item())
print("Norm of 1st layer for modified_deep_nn:", torch.norm(modified_nn_deep.features[0].weight).item())

# Initialize a modified lightweight network with the same seed as our other lightweight instances. This will be trained from scratch to examine the effectiveness of cosine loss minimization.
torch.manual_seed(42)
modified_nn_light = ModifiedLightNNCosine(num_classes=10).to(device)
print("Norm of 1st layer:", torch.norm(modified_nn_light.features[0].weight).item())
Norm of 1st layer for deep_nn: 7.5062713623046875
Norm of 1st layer for modified_deep_nn: 7.5062713623046875
Norm of 1st layer: 2.327361822128296

自然地,我们需要更改训练循环,因为现在模型返回一个元组 (logits, hidden_representation)。使用一个示例输入张量,我们可以打印它们的形状。

# Create a sample input tensor
sample_input = torch.randn(128, 3, 32, 32).to(device) # Batch size: 128, Filters: 3, Image size: 32x32

# Pass the input through the student
logits, hidden_representation = modified_nn_light(sample_input)

# Print the shapes of the tensors
print("Student logits shape:", logits.shape) # batch_size x total_classes
print("Student hidden representation shape:", hidden_representation.shape) # batch_size x hidden_representation_size

# Pass the input through the teacher
logits, hidden_representation = modified_nn_deep(sample_input)

# Print the shapes of the tensors
print("Teacher logits shape:", logits.shape) # batch_size x total_classes
print("Teacher hidden representation shape:", hidden_representation.shape) # batch_size x hidden_representation_size
Student logits shape: torch.Size([128, 10])
Student hidden representation shape: torch.Size([128, 1024])
Teacher logits shape: torch.Size([128, 10])
Teacher hidden representation shape: torch.Size([128, 1024])

在我们的案例中,hidden_representation_size1024。这是学生模型最后一个卷积层的扁平化特征图,正如你所见,它是其分类器的输入。对于教师模型来说,它也是 1024,因为我们通过 avg_pool1d2048 进行了调整。这里应用的损失仅影响学生在损失计算之前的权重。换句话说,它不会影响学生的分类器。修改后的训练循环如下:

../_static/img/knowledge_distillation/cosine_loss_distillation.png

在余弦损失最小化中,我们希望通过向学生返回梯度来最大化两个表示的余弦相似度:

def train_cosine_loss(teacher, student, train_loader, epochs, learning_rate, hidden_rep_loss_weight, ce_loss_weight, device):
    ce_loss = nn.CrossEntropyLoss()
    cosine_loss = nn.CosineEmbeddingLoss()
    optimizer = optim.Adam(student.parameters(), lr=learning_rate)

    teacher.to(device)
    student.to(device)
    teacher.eval()  # Teacher set to evaluation mode
    student.train() # Student to train mode

    for epoch in range(epochs):
        running_loss = 0.0
        for inputs, labels in train_loader:
            inputs, labels = inputs.to(device), labels.to(device)

            optimizer.zero_grad()

            # Forward pass with the teacher model and keep only the hidden representation
            with torch.no_grad():
                _, teacher_hidden_representation = teacher(inputs)

            # Forward pass with the student model
            student_logits, student_hidden_representation = student(inputs)

            # Calculate the cosine loss. Target is a vector of ones. From the loss formula above we can see that is the case where loss minimization leads to cosine similarity increase.
            hidden_rep_loss = cosine_loss(student_hidden_representation, teacher_hidden_representation, target=torch.ones(inputs.size(0)).to(device))

            # Calculate the true label loss
            label_loss = ce_loss(student_logits, labels)

            # Weighted sum of the two losses
            loss = hidden_rep_loss_weight * hidden_rep_loss + ce_loss_weight * label_loss

            loss.backward()
            optimizer.step()

            running_loss += loss.item()

        print(f"Epoch {epoch+1}/{epochs}, Loss: {running_loss / len(train_loader)}")

出于同样的原因,我们需要修改我们的测试函数。这里我们忽略了模型返回的隐藏表示。

def test_multiple_outputs(model, test_loader, device):
    model.to(device)
    model.eval()

    correct = 0
    total = 0

    with torch.no_grad():
        for inputs, labels in test_loader:
            inputs, labels = inputs.to(device), labels.to(device)

            outputs, _ = model(inputs) # Disregard the second tensor of the tuple
            _, predicted = torch.max(outputs.data, 1)

            total += labels.size(0)
            correct += (predicted == labels).sum().item()

    accuracy = 100 * correct / total
    print(f"Test Accuracy: {accuracy:.2f}%")
    return accuracy

在这种情况下,我们可以轻松地在同一个函数中同时包含知识蒸馏和余弦损失最小化。在教师-学生范式中,结合方法以实现更好的性能是很常见的。 目前,我们可以运行一个简单的训练-测试会话。

# Train and test the lightweight network with cross entropy loss
train_cosine_loss(teacher=modified_nn_deep, student=modified_nn_light, train_loader=train_loader, epochs=10, learning_rate=0.001, hidden_rep_loss_weight=0.25, ce_loss_weight=0.75, device=device)
test_accuracy_light_ce_and_cosine_loss = test_multiple_outputs(modified_nn_light, test_loader, device)
Epoch 1/10, Loss: 1.3019573956804202
Epoch 2/10, Loss: 1.0648430796230541
Epoch 3/10, Loss: 0.9631246839033063
Epoch 4/10, Loss: 0.8873560082577073
Epoch 5/10, Loss: 0.8324901197877381
Epoch 6/10, Loss: 0.7894727920022462
Epoch 7/10, Loss: 0.7494128462298751
Epoch 8/10, Loss: 0.7148379474649649
Epoch 9/10, Loss: 0.6762727979199051
Epoch 10/10, Loss: 0.6496843107216194
Test Accuracy: 71.17%

中级回归器运行

我们的简单最小化方法并不能保证更好的结果,原因有几个,其中之一是向量的维度。 对于高维向量,余弦相似度通常比欧几里得距离效果更好, 但我们处理的是每个有1024个分量的向量,因此提取有意义的相似性要困难得多。 此外,正如我们提到的,推动教师和学生的隐藏表示匹配并没有理论支持。 我们没有充分的理由去追求这些向量的1:1匹配。 我们将通过引入一个称为回归器的额外网络来提供一个训练干预的最终示例。 目标是首先在卷积层之后提取教师的特征图, 然后在卷积层之后提取学生的特征图,最后尝试匹配这些图。 然而,这次我们将在网络之间引入一个回归器,以促进匹配过程。 回归器将是可训练的,理想情况下会比我们简单的余弦损失最小化方案做得更好。 它的主要工作是匹配这些特征图的维度,以便我们可以在教师和学生之间正确定义一个损失函数。 定义这样的损失函数提供了一个教学的“路径”,这基本上是一个反向传播梯度的流程,将改变学生的权重。 专注于我们原始网络中每个分类器之前的卷积层输出,我们有以下形状:

# Pass the sample input only from the convolutional feature extractor
convolutional_fe_output_student = nn_light.features(sample_input)
convolutional_fe_output_teacher = nn_deep.features(sample_input)

# Print their shapes
print("Student's feature extractor output shape: ", convolutional_fe_output_student.shape)
print("Teacher's feature extractor output shape: ", convolutional_fe_output_teacher.shape)
Student's feature extractor output shape:  torch.Size([128, 16, 8, 8])
Teacher's feature extractor output shape:  torch.Size([128, 32, 8, 8])

我们为教师准备了32个过滤器,为学生准备了16个过滤器。 我们将包含一个可训练的层,该层将学生的特征图转换为教师特征图的形状。 在实践中,我们修改了轻量级类,以返回在匹配卷积特征图大小的中间回归器之后的隐藏状态,并修改教师类以返回最终卷积层的输出,而不进行池化或展平。

../_static/img/knowledge_distillation/fitnets_knowledge_distill.png

可训练层匹配中间张量的形状,并且均方误差(MSE)正确定义:

class ModifiedDeepNNRegressor(nn.Module):
    def __init__(self, num_classes=10):
        super(ModifiedDeepNNRegressor, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 128, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(128, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(64, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(64, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )
        self.classifier = nn.Sequential(
            nn.Linear(2048, 512),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(512, num_classes)
        )

    def forward(self, x):
        x = self.features(x)
        conv_feature_map = x
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x, conv_feature_map

class ModifiedLightNNRegressor(nn.Module):
    def __init__(self, num_classes=10):
        super(ModifiedLightNNRegressor, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            nn.Conv2d(16, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
        )
        # Include an extra regressor (in our case linear)
        self.regressor = nn.Sequential(
            nn.Conv2d(16, 32, kernel_size=3, padding=1)
        )
        self.classifier = nn.Sequential(
            nn.Linear(1024, 256),
            nn.ReLU(),
            nn.Dropout(0.1),
            nn.Linear(256, num_classes)
        )

    def forward(self, x):
        x = self.features(x)
        regressor_output = self.regressor(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x, regressor_output

之后,我们必须再次更新我们的训练循环。这次,我们提取学生的回归器输出,教师的特征图,我们计算这些张量上的MSE(它们具有完全相同的形状,因此定义正确),并基于该损失反向传播梯度,此外还有分类任务的常规交叉熵损失。

def train_mse_loss(teacher, student, train_loader, epochs, learning_rate, feature_map_weight, ce_loss_weight, device):
    ce_loss = nn.CrossEntropyLoss()
    mse_loss = nn.MSELoss()
    optimizer = optim.Adam(student.parameters(), lr=learning_rate)

    teacher.to(device)
    student.to(device)
    teacher.eval()  # Teacher set to evaluation mode
    student.train() # Student to train mode

    for epoch in range(epochs):
        running_loss = 0.0
        for inputs, labels in train_loader:
            inputs, labels = inputs.to(device), labels.to(device)

            optimizer.zero_grad()

            # Again ignore teacher logits
            with torch.no_grad():
                _, teacher_feature_map = teacher(inputs)

            # Forward pass with the student model
            student_logits, regressor_feature_map = student(inputs)

            # Calculate the loss
            hidden_rep_loss = mse_loss(regressor_feature_map, teacher_feature_map)

            # Calculate the true label loss
            label_loss = ce_loss(student_logits, labels)

            # Weighted sum of the two losses
            loss = feature_map_weight * hidden_rep_loss + ce_loss_weight * label_loss

            loss.backward()
            optimizer.step()

            running_loss += loss.item()

        print(f"Epoch {epoch+1}/{epochs}, Loss: {running_loss / len(train_loader)}")

# Notice how our test function remains the same here with the one we used in our previous case. We only care about the actual outputs because we measure accuracy.

# Initialize a ModifiedLightNNRegressor
torch.manual_seed(42)
modified_nn_light_reg = ModifiedLightNNRegressor(num_classes=10).to(device)

# We do not have to train the modified deep network from scratch of course, we just load its weights from the trained instance
modified_nn_deep_reg = ModifiedDeepNNRegressor(num_classes=10).to(device)
modified_nn_deep_reg.load_state_dict(nn_deep.state_dict())

# Train and test once again
train_mse_loss(teacher=modified_nn_deep_reg, student=modified_nn_light_reg, train_loader=train_loader, epochs=10, learning_rate=0.001, feature_map_weight=0.25, ce_loss_weight=0.75, device=device)
test_accuracy_light_ce_and_mse_loss = test_multiple_outputs(modified_nn_light_reg, test_loader, device)
Epoch 1/10, Loss: 1.7312568727966464
Epoch 2/10, Loss: 1.3489013407236474
Epoch 3/10, Loss: 1.2052425062260055
Epoch 4/10, Loss: 1.108028480921255
Epoch 5/10, Loss: 1.028127753673612
Epoch 6/10, Loss: 0.9665506834264301
Epoch 7/10, Loss: 0.9110444729285472
Epoch 8/10, Loss: 0.861484837196672
Epoch 9/10, Loss: 0.8197113380712622
Epoch 10/10, Loss: 0.7801015764246206
Test Accuracy: 71.08%

预计最终方法将比CosineLoss效果更好,因为现在我们在教师和学生之间允许了一个可训练的层, 这给了学生在学习时一些灵活的空间,而不是迫使学生复制教师的表示。 包含额外的网络是基于提示的蒸馏背后的想法。

print(f"Teacher accuracy: {test_accuracy_deep:.2f}%")
print(f"Student accuracy without teacher: {test_accuracy_light_ce:.2f}%")
print(f"Student accuracy with CE + KD: {test_accuracy_light_ce_and_kd:.2f}%")
print(f"Student accuracy with CE + CosineLoss: {test_accuracy_light_ce_and_cosine_loss:.2f}%")
print(f"Student accuracy with CE + RegressorMSE: {test_accuracy_light_ce_and_mse_loss:.2f}%")
Teacher accuracy: 75.45%
Student accuracy without teacher: 70.15%
Student accuracy with CE + KD: 70.75%
Student accuracy with CE + CosineLoss: 71.17%
Student accuracy with CE + RegressorMSE: 71.08%

结论

上述方法均未增加网络的参数数量或推理时间, 因此性能的提升仅需在训练期间计算梯度的微小成本。 在机器学习应用中,我们主要关注推理时间,因为训练发生在模型部署之前。 如果我们的轻量级模型对于部署来说仍然过于沉重,我们可以应用不同的想法,例如训练后量化。 额外的损失可以应用于许多任务,而不仅仅是分类,你可以尝试调整系数、 温度或神经元数量等参数。请随意调整上述教程中的任何数字, 但请记住,如果你改变神经元/过滤器的数量,可能会出现形状不匹配的情况。

欲了解更多信息,请参阅:

脚本总运行时间: (7 分钟 45.085 秒)

Gallery generated by Sphinx-Gallery