1、迁移学习

（抄自CS231n Convolutional Neural Networks for Visual Recognition）

        在实践中，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罕见。相反，通常会在非常大的数据集上预训练 ConvNet（例如 ImageNet，其中包含 120 万张图像和 1000 个类别），然后将 ConvNet 用作初始化或固定特征提取器来执行感兴趣的任务。三个主要的迁移学习方案如下所示：

• ConvNet 作为固定特征提取器。在 ImageNet 上预训练 ConvNet，删除最后一个全连接层（该层的输出是 ImageNet 等不同任务的 1000 个类分数），然后将 ConvNet 的其余部分视为新数据集的固定特征提取器。在 AlexNet 中，这将为每个图像计算一个 4096-D 向量，该图像包含紧接在分类器之前的隐藏层的激活。我们将这些特征称为 CNN 代码。对于性能来说，如果这些代码在 ImageNet 上训练 ConvNet 期间也被阈值化（通常情况如此），那么这些代码是 ReLUd（即阈值为零）是很重要的。提取所有图像的 4096-D 代码后，为新数据集训练线性分类器（例如线性 SVM 或 Softmax 分类器）。
• 微调 ConvNet。第二种策略是，不仅要在新数据集上替换和重新训练ConvNet上的分类器，还要通过继续反向传播来微调预训练网络的权重。可以对 ConvNet 的所有层进行微调，也可以将一些早期的层固定（由于过度拟合问题）并仅微调网络的某些更高级别的部分。这是由于观察到 ConvNet 的早期特征包含更通用的特征（例如边缘检测器或颜色斑点检测器），这些特征应该对许多任务有用，但 ConvNet 的后续层逐渐变得更加特定于原始数据集中包含的类的详细信息。例如，对于包含许多犬种的 ImageNet，ConvNet 的很大一部分表示能力可能专门用于区分犬种的功能。
• 预训练模型。由于现代 ConvNet 需要 2-3 周的时间才能在 ImageNet 上的多个 GPU 上进行训练，因此通常会看到人们发布最终的 ConvNet 检查点，以造福其他可以使用网络进行微调的人。例如，Caffe 库有一个模型动物园，人们可以在其中共享他们的网络权重。

 2、数据准备

        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 downloaddataset_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)

3、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。

#定义输入
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 = 32if usage == "train":# Define map operations for training datasettrans = [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 datasettrans = [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_setdataset_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()

4、数据集可视化

        从mindspore.dataset.ImageFolderDataset接口中加载的训练数据集返回值为字典，用户可通过 create_dict_iterator 接口创建数据迭代器，使用 next 迭代访问数据集。本章中 batch_size 设为18，所以使用 next 一次可获取18个图像及标签数据。

data = next(dataset_train.create_dict_iterator())
images = data["image"]
labels = data["label"]print("Tensor of image", images.shape)
print("Labels:", labels)

4.1 图像标签可视化 

 对获取到的图像及标签数据进行可视化，标题为图像对应的label名称。

import matplotlib.pyplot as plt
import numpy as np# class_name对应label，按文件夹字符串从小到大的顺序标记label
class_name = {0: "dogs", 1: "wolves"}plt.figure(figsize=(5, 5))
for i in range(4):# 获取图像及其对应的labeldata_image = images[i].asnumpy()data_label = labels[i]# 处理图像供展示使用data_image = np.transpose(data_image, (1, 2, 0))mean = np.array([0.485, 0.456, 0.406])std = np.array([0.229, 0.224, 0.225])data_image = std * data_image + meandata_image = np.clip(data_image, 0, 1)# 显示图像plt.subplot(2, 2, i+1)plt.imshow(data_image)plt.title(class_name[int(labels[i].asnumpy())])plt.axis("off")plt.show()

 5、训练模型

        使用ResNet50模型进行训练。搭建好模型框架后，通过将pretrained参数设置为True来下载ResNet50的预训练模型并将权重参数加载到网络中。

5.1 构建网络

from typing import Type, Union, List, Optional
from mindspore import nn, train
from mindspore.common.initializer import Normalweight_init = Normal(mean=0, sigma=0.02)
gamma_init = Normal(mean=1, sigma=0.02)

 

class ResidualBlockBase(nn.Cell):expansion: int = 1  # 最后一个卷积核数量与第一个卷积核数量相等def __init__(self, in_channel: int, out_channel: int,stride: int = 1, norm: Optional[nn.Cell] = None,down_sample: Optional[nn.Cell] = None) -> None:super(ResidualBlockBase, self).__init__()if not norm:self.norm = nn.BatchNorm2d(out_channel)else:self.norm = normself.conv1 = nn.Conv2d(in_channel, out_channel,kernel_size=3, stride=stride,weight_init=weight_init)self.conv2 = nn.Conv2d(in_channel, out_channel,kernel_size=3, weight_init=weight_init)self.relu = nn.ReLU()self.down_sample = down_sampledef construct(self, x):"""ResidualBlockBase construct."""identity = x  # shortcuts分支out = self.conv1(x)  # 主分支第一层：3*3卷积层out = self.norm(out)out = self.relu(out)out = self.conv2(out)  # 主分支第二层：3*3卷积层out = self.norm(out)if self.down_sample is not None:identity = self.down_sample(x)out += identity  # 输出为主分支与shortcuts之和out = self.relu(out)return out

class ResidualBlock(nn.Cell):expansion = 4  # 最后一个卷积核的数量是第一个卷积核数量的4倍def __init__(self, in_channel: int, out_channel: int,stride: int = 1, down_sample: Optional[nn.Cell] = None) -> None:super(ResidualBlock, self).__init__()self.conv1 = nn.Conv2d(in_channel, out_channel,kernel_size=1, weight_init=weight_init)self.norm1 = nn.BatchNorm2d(out_channel)self.conv2 = nn.Conv2d(out_channel, out_channel,kernel_size=3, stride=stride,weight_init=weight_init)self.norm2 = nn.BatchNorm2d(out_channel)self.conv3 = nn.Conv2d(out_channel, out_channel * self.expansion,kernel_size=1, weight_init=weight_init)self.norm3 = nn.BatchNorm2d(out_channel * self.expansion)self.relu = nn.ReLU()self.down_sample = down_sampledef construct(self, x):identity = x  # shortscuts分支out = self.conv1(x)  # 主分支第一层：1*1卷积层out = self.norm1(out)out = self.relu(out)out = self.conv2(out)  # 主分支第二层：3*3卷积层out = self.norm2(out)out = self.relu(out)out = self.conv3(out)  # 主分支第三层：1*1卷积层out = self.norm3(out)if self.down_sample is not None:identity = self.down_sample(x)out += identity  # 输出为主分支与shortcuts之和out = self.relu(out)return out

 

def make_layer(last_out_channel, block: Type[Union[ResidualBlockBase, ResidualBlock]],channel: int, block_nums: int, stride: int = 1):down_sample = None  # shortcuts分支if stride != 1 or last_out_channel != channel * block.expansion:down_sample = nn.SequentialCell([nn.Conv2d(last_out_channel, channel * block.expansion,kernel_size=1, stride=stride, weight_init=weight_init),nn.BatchNorm2d(channel * block.expansion, gamma_init=gamma_init)])layers = []layers.append(block(last_out_channel, channel, stride=stride, down_sample=down_sample))in_channel = channel * block.expansion# 堆叠残差网络for _ in range(1, block_nums):layers.append(block(in_channel, channel))return nn.SequentialCell(layers)

from mindspore import load_checkpoint, load_param_into_netclass 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64self.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 xdef _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 modeldef 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)

 5.2 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 timenet_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 lossgrad_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()})

5.3 训练和评估

         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"

import mindspore as ms
import matplotlib.pyplot as plt
import os
import time
# 开始循环训练
print("Start Training Loop ...")best_acc = 0for 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) * 1000step_seconds = epoch_seconds/step_size_trainprint("-" * 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 = accif 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)

 5.4 可视化模型预测

        使用固定特征得到的best.ckpt文件对对验证集的狼和狗图像数据进行预测。若预测字体为蓝色即为预测正确，若预测字体为红色则预测错误。

import matplotlib.pyplot as plt
import mindspore as msdef visualize_model(best_ckpt_path, val_ds):net = resnet50()# 全连接层输入层的大小in_channels = net.fc.in_channels# 输出通道数大小为狼狗分类数2head = nn.Dense(in_channels, 2)# 重置全连接层net.fc = head# 平均池化层kernel size为7avg_pool = nn.AvgPool2d(kernel_size=7)# 重置平均池化层net.avg_pool = avg_pool# 加载模型参数param_dict = ms.load_checkpoint(best_ckpt_path)ms.load_param_into_net(net, param_dict)model = train.Model(net)# 加载验证集的数据进行验证data = next(val_ds.create_dict_iterator())images = data["image"].asnumpy()labels = data["label"].asnumpy()class_name = {0: "dogs", 1: "wolves"}# 预测图像类别output = model.predict(ms.Tensor(data['image']))pred = np.argmax(output.asnumpy(), axis=1)# 显示图像及图像的预测值plt.figure(figsize=(5, 5))for i in range(4):plt.subplot(2, 2, i + 1)# 若预测正确，显示为蓝色；若预测错误，显示为红色color = 'blue' if pred[i] == labels[i] else 'red'plt.title('predict:{}'.format(class_name[pred[i]]), color=color)picture_show = np.transpose(images[i], (1, 2, 0))mean = np.array([0.485, 0.456, 0.406])std = np.array([0.229, 0.224, 0.225])picture_show = std * picture_show + meanpicture_show = np.clip(picture_show, 0, 1)plt.imshow(picture_show)plt.axis('off')plt.show()

visualize_model(best_ckpt_path, dataset_val)