以下为官方活动的学习笔记兼打卡记录，大部分内容来自活动资料，稍有删改，内含跳转至MindSpore文档的超链接，可作为字典查询。

ResNet50迁移学习

在实际应用场景中，由于训练数据集不足，所以很少有人会从头开始训练整个网络。普遍的做法是，在一个非常大的基础数据集上训练得到一个预训练模型，然后使用该模型来初始化网络的权重参数或作为固定特征提取器应用于特定的任务中。本章将使用迁移学习的方法对ImageNet数据集中的狼和狗图像进行分类。

迁移学习详细内容见Stanford University CS231n。

一、数据准备

1.1 下载数据集

下载案例所用到的狗与狼分类数据集，数据集中的图像来自于ImageNet，每个分类有大约120张训练图像与30张验证图像。使用download接口下载数据集，并将下载后的数据集自动解压到当前目录下。

from download import download

dataset_url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/intermediate/Canidae_data.zip"

download(dataset_url, "./datasets-Canidae", kind="zip", replace=True)

数据集目录结构如下：

datasets-Canidae/data/
└── Canidae
    ├── train
    │   ├── dogs
    │   └── wolves
    └── val
        ├── dogs
        └── wolves

1.2 加载数据集

狼狗数据集提取自ImageNet分类数据集，使用mindspore.dataset.ImageFolderDataset接口来加载数据集，并进行相关图像增强操作。

batch_size = 18                             # 批量大小
image_size = 224                            # 训练图像空间大小
num_epochs = 5                             # 训练周期数
lr = 0.001                                  # 学习率
momentum = 0.9                              # 动量
workers = 4                                 # 并行线程个数

import mindspore as ms
import mindspore.dataset as ds
import mindspore.dataset.vision as vision

# 数据集目录路径
data_path_train = "./datasets-Canidae/data/Canidae/train/"
data_path_val = "./datasets-Canidae/data/Canidae/val/"

# 创建训练数据集
def create_dataset_canidae(dataset_path, usage):
    """数据加载"""
    data_set = ds.ImageFolderDataset(dataset_path,
                                     num_parallel_workers=workers,
                                     shuffle=True,)

    # 数据增强操作
    mean = [0.485 * 255, 0.456 * 255, 0.406 * 255]
    std = [0.229 * 255, 0.224 * 255, 0.225 * 255]
    scale = 32

    if usage == "train":
        # Define map operations for training dataset
        trans = [
            vision.RandomCropDecodeResize(size=image_size, scale=(0.08, 1.0), ratio=(0.75, 1.333)),
            vision.RandomHorizontalFlip(prob=0.5),
            vision.Normalize(mean=mean, std=std),
            vision.HWC2CHW()
        ]
    else:
        # Define map operations for inference dataset
        trans = [
            vision.Decode(),
            vision.Resize(image_size + scale),
            vision.CenterCrop(image_size),
            vision.Normalize(mean=mean, std=std),
            vision.HWC2CHW()
        ]


    # 数据映射操作
    data_set = data_set.map(
        operations=trans,
        input_columns='image',
        num_parallel_workers=workers)


    # 批量操作
    data_set = data_set.batch(batch_size)

    return data_set


dataset_train = create_dataset_canidae(data_path_train, "train")
step_size_train = dataset_train.get_dataset_size()

dataset_val = create_dataset_canidae(data_path_val, "val")
step_size_val = dataset_val.get_dataset_size()

1.3 数据集可视化

可以看到目前训练集中的图像呈现了数据增强的效果。

二、训练模型

2.1 构建Resnet50网络

具体代码可参考第六天的学习记录。

class ResNet(nn.Cell):
    def __init__(self, block: Type[Union[ResidualBlockBase, ResidualBlock]],
                 layer_nums: List[int], num_classes: int, input_channel: int) -> None:
        super(ResNet, self).__init__()

        self.relu = nn.ReLU()
        # 第一个卷积层，输入channel为3（彩色图像），输出channel为64
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, weight_init=weight_init)
        self.norm = nn.BatchNorm2d(64)
        # 最大池化层，缩小图片的尺寸
        self.max_pool = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode='same')
        # 各个残差网络结构块定义，
        self.layer1 = make_layer(64, block, 64, layer_nums[0])
        self.layer2 = make_layer(64 * block.expansion, block, 128, layer_nums[1], stride=2)
        self.layer3 = make_layer(128 * block.expansion, block, 256, layer_nums[2], stride=2)
        self.layer4 = make_layer(256 * block.expansion, block, 512, layer_nums[3], stride=2)
        # 平均池化层
        self.avg_pool = nn.AvgPool2d()
        # flattern层
        self.flatten = nn.Flatten()
        # 全连接层
        self.fc = nn.Dense(in_channels=input_channel, out_channels=num_classes)

    def construct(self, x):

        x = self.conv1(x)
        x = self.norm(x)
        x = self.relu(x)
        x = self.max_pool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avg_pool(x)
        x = self.flatten(x)
        x = self.fc(x)

        return x


def _resnet(model_url: str,
            block: Type[Union[ResidualBlockBase, ResidualBlock]],
            layers: List[int],
            num_classes: int,
            pretrained: bool,
            pretrianed_ckpt: str,
            input_channel: int):
    model = ResNet(block, layers, num_classes, input_channel)

    if pretrained:
        # 加载预训练模型
        download(url=model_url, path=pretrianed_ckpt, replace=True)
        param_dict = load_checkpoint(pretrianed_ckpt)
        load_param_into_net(model, param_dict)

    return model


def resnet50(num_classes: int = 1000, pretrained: bool = False):
    "ResNet50模型"
    resnet50_url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/models/application/resnet50_224_new.ckpt"
    resnet50_ckpt = "./LoadPretrainedModel/resnet50_224_new.ckpt"
    return _resnet(resnet50_url, ResidualBlock, [3, 4, 6, 3], num_classes,
                   pretrained, resnet50_ckpt, 2048)

2.2 固定特征进行训练

使用固定特征进行训练的时候，需要冻结除最后一层之外的所有网络层。通过设置 requires_grad == False 冻结参数，以便不在反向传播中计算梯度。

import mindspore as ms
import matplotlib.pyplot as plt
import os
import time

net_work = resnet50(pretrained=True)

# 全连接层输入层的大小
in_channels = net_work.fc.in_channels
# 输出通道数大小为狼狗分类数2
head = nn.Dense(in_channels, 2)
# 重置全连接层
net_work.fc = head

# 平均池化层kernel size为7
avg_pool = nn.AvgPool2d(kernel_size=7)
# 重置平均池化层
net_work.avg_pool = avg_pool

# 冻结除最后一层外的所有参数
for param in net_work.get_parameters():
    if param.name not in ["fc.weight", "fc.bias"]:
        param.requires_grad = False

# 定义优化器和损失函数
opt = nn.Momentum(params=net_work.trainable_params(), learning_rate=lr, momentum=0.5)
loss_fn = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')


def forward_fn(inputs, targets):
    logits = net_work(inputs)
    loss = loss_fn(logits, targets)

    return loss

grad_fn = ms.value_and_grad(forward_fn, None, opt.parameters)

def train_step(inputs, targets):
    loss, grads = grad_fn(inputs, targets)
    opt(grads)
    return loss

# 实例化模型
model1 = train.Model(net_work, loss_fn, opt, metrics={"Accuracy": train.Accuracy()})

2.3 训练与评估

开始训练模型，与没有预训练模型相比，将节约一大半时间，因为此时可以不用计算部分梯度。保存评估精度最高的ckpt文件于当前路径的./BestCheckpoint/resnet50-best-freezing-param.ckpt。

import mindspore as ms
import matplotlib.pyplot as plt
import os
import time
dataset_train = create_dataset_canidae(data_path_train, "train")
step_size_train = dataset_train.get_dataset_size()

dataset_val = create_dataset_canidae(data_path_val, "val")
step_size_val = dataset_val.get_dataset_size()

num_epochs = 5

# 创建迭代器
data_loader_train = dataset_train.create_tuple_iterator(num_epochs=num_epochs)
data_loader_val = dataset_val.create_tuple_iterator(num_epochs=num_epochs)
best_ckpt_dir = "./BestCheckpoint"
best_ckpt_path = "./BestCheckpoint/resnet50-best-freezing-param.ckpt"

# 开始循环训练
print("Start Training Loop ...")

best_acc = 0

for epoch in range(num_epochs):
    losses = []
    net_work.set_train()

    epoch_start = time.time()

    # 为每轮训练读入数据
    for i, (images, labels) in enumerate(data_loader_train):
        labels = labels.astype(ms.int32)
        loss = train_step(images, labels)
        losses.append(loss)

    # 每个epoch结束后，验证准确率

    acc = model1.eval(dataset_val)['Accuracy']

    epoch_end = time.time()
    epoch_seconds = (epoch_end - epoch_start) * 1000
    step_seconds = epoch_seconds/step_size_train

    print("-" * 20)
    print("Epoch: [%3d/%3d], Average Train Loss: [%5.3f], Accuracy: [%5.3f]" % (
        epoch+1, num_epochs, sum(losses)/len(losses), acc
    ))
    print("epoch time: %5.3f ms, per step time: %5.3f ms" % (
        epoch_seconds, step_seconds
    ))

    if acc > best_acc:
        best_acc = acc
        if not os.path.exists(best_ckpt_dir):
            os.mkdir(best_ckpt_dir)
        ms.save_checkpoint(net_work, best_ckpt_path)

print("=" * 80)
print(f"End of validation the best Accuracy is: {best_acc: 5.3f}, "
      f"save the best ckpt file in {best_ckpt_path}", flush=True)

2.4 预测结果可视化

学习记录