PyTorch深度学习实践：从分类到生成-PyTorch视觉迁移学习

计算机视觉迁移学习

学习目标

在本课程中，将学习如何使用迁移学习来训练卷积神经网络进行图像分类。

相关知识点

• 微调卷积网络

学习内容

1 微调卷积网络

实际上，很少有人从头开始训练整个卷积网络（使用随机初始化），因为拥有足够大小的数据集相对较少。相反，通常会在一个非常大的数据集（例如包含 120 万张图像和 1000 个类别的 ImageNet）上预训练一个卷积网络，然后将该卷积网络用作感兴趣任务的初始化或固定特征提取器。

这两种主要的迁移学习场景如下：

• 微调卷积网络：不是随机初始化，而是用预训练的网络（例如在 ImageNet 1000 数据集上训练的网络）来初始化网络。其余的训练过程与通常的训练过程相同。
• 卷积网络作为固定特征提取器：在这种情况下，除了最后一个全连接层外，我们将冻结网络的所有权重。最后一个全连接层将被一个具有随机权重的新层替换，并且只训练这一层。

%matplotlib inline

from __future__ import print_function, division

import torch
import torch_npu
from torch_npu.contrib import transfer_to_npu
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import torch.backends.cudnn as cudnn
import numpy as np
import torchvision
from torchvision import datasets, models, transforms
import matplotlib.pyplot as plt
import time
import os
import copy

cudnn.benchmark = True
plt.ion()

1.1 加载数据

我们将使用 torchvision 和 torch.utils.data 包来加载数据。

今天我们要解决的问题是训练一个模型来对蚂蚁和蜜蜂进行分类。我们有大约 120 张蚂蚁和蜜蜂的训练图像，每个类别各有 75 张验证图像。通常情况下，这是一个非常小的数据集，如果从头开始训练，很难泛化。由于我们使用了迁移学习，我们应该能够合理地泛化。

这个数据集是 ImageNet 的一个非常小的子集。

!wget https://model-community-picture.obs.cn-north-4.myhuaweicloud.com/ascend-zone/notebook_datasets/cfeec85ee9df11efac8afa163edcddae/hymenoptera_data.zip
!unzip hymenoptera_data.zip

data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'val': transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
}

data_dir = 'hymenoptera_data'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
                                          data_transforms[x])
                  for x in ['train', 'val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4,
                                             shuffle=True, num_workers=4)
              for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_names = image_datasets['train'].classes

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

1.2 可视化一些图像

让我们可视化一些训练图像，以便理解数据增强。

def imshow(inp, title=None):
    """Imshow for Tensor."""
    inp = inp.numpy().transpose((1, 2, 0))
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.225])
    inp = std * inp + mean
    inp = np.clip(inp, 0, 1)
    plt.imshow(inp)
    if title is not None:
        plt.title(title)
    plt.pause(0.001)


inputs, classes = next(iter(dataloaders['train']))

out = torchvision.utils.make_grid(inputs)

imshow(out, title=[class_names[x] for x in classes])

1.3 训练模型

现在，我们来编写一个通用的函数来训练模型。在这里，我们将展示：

• 调整学习率
• 保存最佳模型

在下面的内容中，参数 scheduler 是来自 torch.optim.lr_scheduler 的一个学习率调度器对象。

def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
    since = time.time()

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0

    for epoch in range(num_epochs):
        print(f'Epoch {epoch}/{num_epochs - 1}')
        print('-' * 10)

        # Each epoch has a training and validation phase
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # Set model to training mode
            else:
                model.eval()   # Set model to evaluate mode

            running_loss = 0.0
            running_corrects = 0

            # Iterate over data.
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)

                    # backward + optimize only if in training phase
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # statistics
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)
            if phase == 'train':
                scheduler.step()

            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]

            print(f'{phase} Loss: {epoch_loss:.4f} Acc: {epoch_acc:.4f}')

            # deep copy the model
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())

        print()

    time_elapsed = time.time() - since
    print(f'Training complete in {time_elapsed // 60:.0f}m {time_elapsed % 60:.0f}s')
    print(f'Best val Acc: {best_acc:4f}')

    # load best model weights
    model.load_state_dict(best_model_wts)
    return model

• Or MPS用户（MacBook pro2019 亲测可用）
  上述NPU适配代码可以使用下列代码替换：

import time
import copy
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import torchvision
from torchvision import datasets, models, transforms

# 1. 首先定义设备（确保设备选择逻辑正确）
device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
print(f"Using device: {device}")

# 2. 修复后的训练函数
def train_model(model, criterion, optimizer, scheduler, dataloaders, dataset_sizes, num_epochs=25):
    since = time.time()
    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0

    for epoch in range(num_epochs):
        print(f'Epoch {epoch}/{num_epochs - 1}')
        print('-' * 10)

        # 每个epoch包含训练和验证阶段
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # 训练模式
            else:
                model.eval()   # 验证模式

            running_loss = 0.0
            running_corrects = 0

            # 迭代数据
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # 清零梯度
                optimizer.zero_grad()

                # 前向传播（仅训练阶段跟踪梯度）
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)

                    # 反向传播+优化（仅训练阶段）
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # 统计指标（关键修复：移除double，改用float）
                running_loss += loss.item() * inputs.size(0)
                # 修复1：将preds和labels的比较结果转为float（MPS支持）
                running_corrects += torch.sum(preds == labels.data).float()

            if phase == 'train':
                scheduler.step()

            # 修复2：使用float计算准确率，避免double
            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects / dataset_sizes[phase]  # 已为float类型

            print(f'{phase} Loss: {epoch_loss:.4f} Acc: {epoch_acc:.4f}')

            # 保存最佳模型权重
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())

        print()

    time_elapsed = time.time() - since
    print(f'Training complete in {time_elapsed // 60:.0f}m {time_elapsed % 60:.0f}s')
    print(f'Best val Acc: {best_acc:.4f}')

    # 加载最佳模型权重
    model.load_state_dict(best_model_wts)
    return model

# 3. 模型初始化（修复pretrained警告）
# 替换deprecated的pretrained参数，使用weights参数
model_ft = models.resnet18(weights=torchvision.models.ResNet18_Weights.IMAGENET1K_V1)
num_ftrs = model_ft.fc.in_features
# 替换全连接层（适配2分类）
model_ft.fc = nn.Linear(num_ftrs, 2)
model_ft = model_ft.to(device)

# 4. 损失函数、优化器、学习率调度器
criterion = nn.CrossEntropyLoss()
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)

# 5. 数据加载逻辑
# 请替换为你的数据集路径
data_dir = "./hymenoptera_data"  # 例如："./hymenoptera_data"
data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'val': transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
}

image_datasets = {
    x: datasets.ImageFolder(f"{data_dir}/{x}", data_transforms[x])
    for x in ['train', 'val']
}
dataloaders = {
    x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4, shuffle=True, num_workers=2)
    for x in ['train', 'val']
}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_names = image_datasets['train'].classes

# 6. 启动训练
model_ft = train_model(
    model_ft, criterion, optimizer_ft, exp_lr_scheduler,
    dataloaders, dataset_sizes, num_epochs=25
)

MPS Out（16分钟，没有NPU快）：

Epoch 0/24
----------
train Loss: 0.5657 Acc: 0.7295
val Loss: 0.2148 Acc: 0.9150

Epoch 1/24
----------
train Loss: 0.5380 Acc: 0.7951
val Loss: 0.3365 Acc: 0.8758

Epoch 2/24
----------
train Loss: 0.5078 Acc: 0.8115
val Loss: 0.3787 Acc: 0.8431

Epoch 3/24
----------
train Loss: 0.5044 Acc: 0.7459
val Loss: 0.3706 Acc: 0.8758

Epoch 4/24
----------
train Loss: 0.5988 Acc: 0.7377
val Loss: 0.6115 Acc: 0.7516

Epoch 5/24
----------
train Loss: 0.5791 Acc: 0.7828
val Loss: 0.2452 Acc: 0.9216

Epoch 6/24
----------
train Loss: 0.4556 Acc: 0.8156
val Loss: 0.2754 Acc: 0.8824

Epoch 7/24
----------
train Loss: 0.2597 Acc: 0.8893
val Loss: 0.2213 Acc: 0.9085

Epoch 8/24
----------
train Loss: 0.3767 Acc: 0.8443
val Loss: 0.2365 Acc: 0.9085

Epoch 9/24
----------
train Loss: 0.3373 Acc: 0.8484
val Loss: 0.2259 Acc: 0.9150

Epoch 10/24
----------
train Loss: 0.2736 Acc: 0.8934
val Loss: 0.2352 Acc: 0.9085

Epoch 11/24
----------
train Loss: 0.2941 Acc: 0.8648
val Loss: 0.2384 Acc: 0.9150

Epoch 12/24
----------
train Loss: 0.2914 Acc: 0.8934
val Loss: 0.2210 Acc: 0.9216

Epoch 13/24
----------
train Loss: 0.2116 Acc: 0.9344
val Loss: 0.2319 Acc: 0.8889

Epoch 14/24
----------
train Loss: 0.2901 Acc: 0.8689
val Loss: 0.2156 Acc: 0.9346

Epoch 15/24
----------
train Loss: 0.2977 Acc: 0.8811
val Loss: 0.2336 Acc: 0.8954

Epoch 16/24
----------
train Loss: 0.2514 Acc: 0.8893
val Loss: 0.2313 Acc: 0.8889

Epoch 17/24
----------
train Loss: 0.2752 Acc: 0.8730
val Loss: 0.2178 Acc: 0.8889

Epoch 18/24
----------
train Loss: 0.2320 Acc: 0.8852
val Loss: 0.2125 Acc: 0.9150

Epoch 19/24
----------
train Loss: 0.2762 Acc: 0.8975
val Loss: 0.2148 Acc: 0.9085

Epoch 20/24
----------
train Loss: 0.2994 Acc: 0.8730
val Loss: 0.2564 Acc: 0.8824

Epoch 21/24
----------
train Loss: 0.2786 Acc: 0.8689
val Loss: 0.2067 Acc: 0.9281

Epoch 22/24
----------
train Loss: 0.2399 Acc: 0.8975
val Loss: 0.2228 Acc: 0.9216

Epoch 23/24
----------
train Loss: 0.2450 Acc: 0.8934
val Loss: 0.2361 Acc: 0.9216

Epoch 24/24
----------
train Loss: 0.2812 Acc: 0.8893
val Loss: 0.2102 Acc: 0.9477

Training complete in 16m 4s
Best val Acc: 0.9477

1.4 可视化模型预测

用于显示一些图像的预测结果的通用函数

def visualize_model(model, num_images=6):
    was_training = model.training
    model.eval()
    images_so_far = 0
    fig = plt.figure()

    with torch.no_grad():
        for i, (inputs, labels) in enumerate(dataloaders['val']):
            inputs = inputs.to(device)
            labels = labels.to(device)

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

            for j in range(inputs.size()[0]):
                images_so_far += 1
                ax = plt.subplot(num_images//2, 2, images_so_far)
                ax.axis('off')
                ax.set_title(f'predicted: {class_names[preds[j]]}')
                imshow(inputs.cpu().data[j])

                if images_so_far == num_images:
                    model.train(mode=was_training)
                    return
        model.train(mode=was_training)

1.5 微调卷积神经网络

加载预训练模型并重置最终的全连接层。

model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
# Here the size of each output sample is set to 2.
# Alternatively, it can be generalized to nn.Linear(num_ftrs, len(class_names)).
model_ft.fc = nn.Linear(num_ftrs, 2)

model_ft = model_ft.to(device)

criterion = nn.CrossEntropyLoss()

# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)

# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)

1.6 训练和评估

在 CPU 上大约需要 15-25 分钟,在MPS版的GPU上大概16分钟。然而，在 NPU 上，它不到一分钟就能完成。

model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,
                       num_epochs=25)

Out：

Epoch 1/24
----------
train Loss: 0.5648 Acc: 0.7541
val Loss: 0.3638 Acc: 0.8497

Epoch 2/24
----------
train Loss: 0.5923 Acc: 0.7500
val Loss: 0.4437 Acc: 0.8627

Epoch 3/24
----------
train Loss: 0.5490 Acc: 0.8115
val Loss: 0.4247 Acc: 0.8627

Epoch 4/24
----------
train Loss: 0.3876 Acc: 0.8320
val Loss: 0.3033 Acc: 0.8824

Epoch 5/24
----------
train Loss: 0.4958 Acc: 0.8156
val Loss: 0.4198 Acc: 0.8431

Epoch 6/24
----------
train Loss: 0.4787 Acc: 0.7992
val Loss: 0.3056 Acc: 0.8693

Epoch 7/24
----------
train Loss: 0.4151 Acc: 0.8074
val Loss: 0.2482 Acc: 0.8889

Epoch 8/24
----------
train Loss: 0.2557 Acc: 0.8975
val Loss: 0.2401 Acc: 0.9150

Epoch 9/24
----------
train Loss: 0.3052 Acc: 0.8689
val Loss: 0.2226 Acc: 0.9150

Epoch 10/24
----------
train Loss: 0.3878 Acc: 0.8197
val Loss: 0.2191 Acc: 0.9216

Epoch 11/24
----------
train Loss: 0.2919 Acc: 0.8770
val Loss: 0.2011 Acc: 0.9281

Epoch 12/24
----------
train Loss: 0.3005 Acc: 0.8525
val Loss: 0.2325 Acc: 0.9150

Epoch 13/24
----------
train Loss: 0.2389 Acc: 0.9139
val Loss: 0.2034 Acc: 0.9216

Epoch 14/24
----------
train Loss: 0.2929 Acc: 0.8648
val Loss: 0.2273 Acc: 0.9020

Epoch 15/24
----------
train Loss: 0.2133 Acc: 0.9139
val Loss: 0.2063 Acc: 0.9085

Epoch 16/24
----------
train Loss: 0.2147 Acc: 0.9057
val Loss: 0.2599 Acc: 0.8693

Epoch 17/24
----------
train Loss: 0.2193 Acc: 0.8975
val Loss: 0.1925 Acc: 0.9412

Epoch 18/24
----------
train Loss: 0.2253 Acc: 0.9098
val Loss: 0.2086 Acc: 0.9216

Epoch 19/24
----------
train Loss: 0.2902 Acc: 0.8811
val Loss: 0.2403 Acc: 0.8758

Epoch 20/24
----------
train Loss: 0.3423 Acc: 0.8484
val Loss: 0.1919 Acc: 0.9281

Epoch 21/24
----------
train Loss: 0.2449 Acc: 0.9098
val Loss: 0.2010 Acc: 0.9542

Epoch 22/24
----------
train Loss: 0.2822 Acc: 0.8689
val Loss: 0.2146 Acc: 0.9085

Epoch 23/24
----------
train Loss: 0.2740 Acc: 0.8648
val Loss: 0.1931 Acc: 0.9216

Epoch 24/24
----------
train Loss: 0.2484 Acc: 0.8893
val Loss: 0.1944 Acc: 0.9216

Training complete in 1m 22s
Best val Acc: 0.954248

visualize_model(model_ft)

1.7 卷积神经网络作为固定特征提取器

在这里，我们需要冻结除最后一层之外的所有网络。我们需要将 requires_grad = False 设置为冻结参数，这样在 backward() 中就不会计算梯度。

model_conv = torchvision.models.resnet18(pretrained=True)
for param in model_conv.parameters():
    param.requires_grad = False

# Parameters of newly constructed modules have requires_grad=True by default
num_ftrs = model_conv.fc.in_features
model_conv.fc = nn.Linear(num_ftrs, 2)

model_conv = model_conv.to(device)

criterion = nn.CrossEntropyLoss()

# Observe that only parameters of final layer are being optimized as
# opposed to before.
optimizer_conv = optim.SGD(model_conv.fc.parameters(), lr=0.001, momentum=0.9)

# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)

• OR MPS版

# 卷积神经网络作为固定特征提取器（适配MPS）
# 替换废弃的pretrained参数，使用weights（torchvision 0.13+规范）
model_conv = models.resnet18(weights=torchvision.models.ResNet18_Weights.IMAGENET1K_V1)

# 冻结除最后一层外的所有参数（MPS兼容）
for param in model_conv.parameters():
    param.requires_grad = False  # 冻结参数，不计算梯度

# 替换最后一层全连接层（新层默认requires_grad=True）
num_ftrs = model_conv.fc.in_features
model_conv.fc = nn.Linear(num_ftrs, 2)  # 2分类任务

# 将模型移至MPS设备（显式指定float32，避免隐式类型问题）
model_conv = model_conv.to(device, dtype=torch.float32)

# 5. 定义损失函数、优化器、学习率调度器
criterion = nn.CrossEntropyLoss()

# 仅优化最后一层的参数（固定特征提取器核心逻辑）
optimizer_conv = optim.SGD(model_conv.fc.parameters(), lr=0.001, momentum=0.9)

# 学习率调度器
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)

# 6. 启动训练（传入dataloaders和dataset_sizes）
model_conv = train_model(
    model_conv, criterion, optimizer_conv, exp_lr_scheduler,
    dataloaders, dataset_sizes, num_epochs=25
)

• MPS Out（12分钟）：

Epoch 0/24
----------
train Loss: 0.6667 Acc: 0.6557
val Loss: 0.4189 Acc: 0.7778

Epoch 1/24
----------
train Loss: 0.4929 Acc: 0.7623
val Loss: 0.2065 Acc: 0.9216

Epoch 2/24
----------
train Loss: 0.4790 Acc: 0.8115
val Loss: 0.1632 Acc: 0.9346

Epoch 3/24
----------
train Loss: 0.4182 Acc: 0.8156
val Loss: 0.2658 Acc: 0.9085

Epoch 4/24
----------
train Loss: 0.3356 Acc: 0.8730
val Loss: 0.1907 Acc: 0.9412

Epoch 5/24
----------
train Loss: 0.5482 Acc: 0.7541
val Loss: 0.1722 Acc: 0.9412

Epoch 6/24
----------
train Loss: 0.5254 Acc: 0.7910
val Loss: 0.2860 Acc: 0.9085

Epoch 7/24
----------
train Loss: 0.4591 Acc: 0.7910
val Loss: 0.1933 Acc: 0.9281

Epoch 8/24
----------
train Loss: 0.2642 Acc: 0.9098
val Loss: 0.1796 Acc: 0.9477

Epoch 9/24
----------
train Loss: 0.3640 Acc: 0.8607
val Loss: 0.1914 Acc: 0.9412

Epoch 10/24
----------
train Loss: 0.3379 Acc: 0.8566
val Loss: 0.1941 Acc: 0.9412

Epoch 11/24
----------
train Loss: 0.3870 Acc: 0.8279
val Loss: 0.1668 Acc: 0.9346

Epoch 12/24
----------
train Loss: 0.3284 Acc: 0.8648
val Loss: 0.1938 Acc: 0.9412

Epoch 13/24
----------
train Loss: 0.4206 Acc: 0.8361
val Loss: 0.1821 Acc: 0.9346

Epoch 14/24
----------
train Loss: 0.3670 Acc: 0.8197
val Loss: 0.1853 Acc: 0.9346

Epoch 15/24
----------
train Loss: 0.3325 Acc: 0.8648
val Loss: 0.1843 Acc: 0.9346

Epoch 16/24
----------
train Loss: 0.3002 Acc: 0.8730
val Loss: 0.1861 Acc: 0.9412

Epoch 17/24
----------
train Loss: 0.2989 Acc: 0.8689
val Loss: 0.1873 Acc: 0.9346

Epoch 18/24
----------
train Loss: 0.4518 Acc: 0.7910
val Loss: 0.2449 Acc: 0.9150

Epoch 19/24
----------
train Loss: 0.3168 Acc: 0.8566
val Loss: 0.1938 Acc: 0.9412

Epoch 20/24
----------
train Loss: 0.3449 Acc: 0.8525
val Loss: 0.1746 Acc: 0.9477

Epoch 21/24
----------
train Loss: 0.3623 Acc: 0.8443
val Loss: 0.2040 Acc: 0.9346

Epoch 22/24
----------
train Loss: 0.3273 Acc: 0.8484
val Loss: 0.1837 Acc: 0.9346

Epoch 23/24
----------
train Loss: 0.3368 Acc: 0.8607
val Loss: 0.1958 Acc: 0.9412

Epoch 24/24
----------
train Loss: 0.3086 Acc: 0.8648
val Loss: 0.1927 Acc: 0.9346

Training complete in 12m 31s
Best val Acc: 0.9477

1.8 训练和评估

在 CPU 上，这将比之前的情况大约节省一半的时间。这是可以预期的，因为对于网络的大部分神经元，不需要计算梯度。然而，正向传播仍然需要计算。

model_conv = train_model(model_conv, criterion, optimizer_conv,
                         exp_lr_scheduler, num_epochs=25)

NPU Out（1分钟）

Epoch 0/24
----------
train Loss: 0.6476 Acc: 0.6352
val Loss: 0.1929 Acc: 0.9412

Epoch 1/24
----------
train Loss: 0.3520 Acc: 0.8525
val Loss: 0.1906 Acc: 0.9281

Epoch 2/24
----------
train Loss: 0.4690 Acc: 0.7746
val Loss: 0.1899 Acc: 0.9477

Epoch 3/24
----------
train Loss: 0.3424 Acc: 0.8361
val Loss: 0.1966 Acc: 0.9346

Epoch 4/24
----------
train Loss: 0.4996 Acc: 0.7910
val Loss: 0.1938 Acc: 0.9412

Epoch 5/24
----------
train Loss: 0.4661 Acc: 0.8033
val Loss: 0.1946 Acc: 0.9346

Epoch 6/24
----------
train Loss: 0.5912 Acc: 0.7951
val Loss: 0.2231 Acc: 0.9412

Epoch 7/24
----------
train Loss: 0.3492 Acc: 0.8443
val Loss: 0.1950 Acc: 0.9477

Epoch 8/24
----------
train Loss: 0.3643 Acc: 0.8156
val Loss: 0.1981 Acc: 0.9477

Epoch 9/24
----------
train Loss: 0.3917 Acc: 0.8115
val Loss: 0.1998 Acc: 0.9412

Epoch 10/24
----------
train Loss: 0.3665 Acc: 0.8156
val Loss: 0.1992 Acc: 0.9412

Epoch 11/24
----------
train Loss: 0.3153 Acc: 0.8730
val Loss: 0.1975 Acc: 0.9477

Epoch 12/24
----------
train Loss: 0.4131 Acc: 0.8115
val Loss: 0.2335 Acc: 0.9346

Epoch 13/24
----------
train Loss: 0.3618 Acc: 0.8566
val Loss: 0.1945 Acc: 0.9477

Epoch 14/24
----------
train Loss: 0.3167 Acc: 0.8443
val Loss: 0.2070 Acc: 0.9412

Epoch 15/24
----------
train Loss: 0.3206 Acc: 0.8811
val Loss: 0.2130 Acc: 0.9412

Epoch 16/24
----------
train Loss: 0.3715 Acc: 0.8361
val Loss: 0.2355 Acc: 0.9281

Epoch 17/24
----------
train Loss: 0.3498 Acc: 0.8443
val Loss: 0.1933 Acc: 0.9412

Epoch 18/24
----------
train Loss: 0.2917 Acc: 0.8607
val Loss: 0.1950 Acc: 0.9477

Epoch 19/24
----------
train Loss: 0.2853 Acc: 0.8730
val Loss: 0.1892 Acc: 0.9477

Epoch 20/24
----------
train Loss: 0.2961 Acc: 0.8770
val Loss: 0.1999 Acc: 0.9477

Epoch 21/24
----------
train Loss: 0.3339 Acc: 0.8402
val Loss: 0.2348 Acc: 0.9412

Epoch 22/24
----------
train Loss: 0.3689 Acc: 0.8525
val Loss: 0.2257 Acc: 0.9412

Epoch 23/24
----------
train Loss: 0.4128 Acc: 0.7992
val Loss: 0.2317 Acc: 0.9281

Epoch 24/24
----------
train Loss: 0.4125 Acc: 0.8115
val Loss: 0.2207 Acc: 0.9281

Training complete in 1m 0s
Best val Acc: 0.947712

visualize_model(model_conv)

plt.ioff()
plt.show()