ResNet网络介绍
ResNet50网络是2015年由微软实验室的何恺明提出,获得ILSVRC2015图像分类竞赛第一名。在ResNet网络提出之前,传统的卷积神经网络都是将一系列的卷积层和池化层堆叠得到的,但当网络堆叠到一定深度时,就会出现退化问题。下图是在CIFAR-10数据集上使用56层网络与20层网络训练误差和测试误差图,由图中数据可以看出,56层网络比20层网络训练误差和测试误差更大,随着网络的加深,其误差并没有如预想的一样减小。
ResNet网络提出了残差网络结构(Residual Network)来减轻退化问题,使用ResNet网络可以实现搭建较深的网络结构(突破1000层)。论文中使用ResNet网络在CIFAR-10数据集上的训练误差与测试误差图如下图所示,图中虚线表示训练误差,实线表示测试误差。由图中数据可以看出,ResNet网络层数越深,其训练误差和测试误差越小。
数据集准备与加载
CIFAR-10数据集共有60000张32*32的彩色图像,分为10个类别,每类有6000张图,数据集一共有50000张训练图片和10000张评估图片。首先,如下示例使用download
接口下载并解压,目前仅支持解析二进制版本的CIFAR-10文件(CIFAR-10 binary version)。
from download import download
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/cifar-10-binary.tar.gz"
download(url, "./datasets-cifar10-bin", kind="tar.gz", replace=True)
import mindspore as ms
import mindspore.dataset as ds
import mindspore.dataset.vision as vision
import mindspore.dataset.transforms as transforms
from mindspore import dtype as mstype
data_dir = "./datasets-cifar10-bin/cifar-10-batches-bin" # 数据集根目录
batch_size = 256 # 批量大小
image_size = 32 # 训练图像空间大小
workers = 4 # 并行线程个数
num_classes = 10 # 分类数量
def create_dataset_cifar10(dataset_dir, usage, resize, batch_size, workers):
data_set = ds.Cifar10Dataset(dataset_dir=dataset_dir,
usage=usage,
num_parallel_workers=workers,
shuffle=True)
trans = []
if usage == "train":
trans += [
vision.RandomCrop((32, 32), (4, 4, 4, 4)),
vision.RandomHorizontalFlip(prob=0.5)
]
trans += [
vision.Resize(resize),
vision.Rescale(1.0 / 255.0, 0.0),
vision.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]),
vision.HWC2CHW()
]
target_trans = transforms.TypeCast(mstype.int32)
# 数据映射操作
data_set = data_set.map(operations=trans,
input_columns='image',
num_parallel_workers=workers)
data_set = data_set.map(operations=target_trans,
input_columns='label',
num_parallel_workers=workers)
# 批量操作
data_set = data_set.batch(batch_size)
return data_set
# 获取处理后的训练与测试数据集
dataset_train = create_dataset_cifar10(dataset_dir=data_dir,
usage="train",
resize=image_size,
batch_size=batch_size,
workers=workers)
step_size_train = dataset_train.get_dataset_size()
dataset_val = create_dataset_cifar10(dataset_dir=data_dir,
usage="test",
resize=image_size,
batch_size=batch_size,
workers=workers)
step_size_val = dataset_val.get_dataset_size()
对CIFAR-10训练数据集进行可视化。
import matplotlib.pyplot as plt
import numpy as np
data_iter = next(dataset_train.create_dict_iterator())
images = data_iter["image"].asnumpy()
labels = data_iter["label"].asnumpy()
print(f"Image shape: {images.shape}, Label shape: {labels.shape}")
# 训练数据集中,前六张图片所对应的标签
print(f"Labels: {labels[:6]}")
classes = []
with open(data_dir + "/batches.meta.txt", "r") as f:
for line in f:
line = line.rstrip()
if line:
classes.append(line)
# 训练数据集的前六张图片
plt.figure()
for i in range(6):
plt.subplot(2, 3, i + 1)
image_trans = np.transpose(images[i], (1, 2, 0))
mean = np.array([0.4914, 0.4822, 0.4465])
std = np.array([0.2023, 0.1994, 0.2010])
image_trans = std * image_trans + mean
image_trans = np.clip(image_trans, 0, 1)
plt.title(f"{classes[labels[i]]}")
plt.imshow(image_trans)
plt.axis("off")
plt.show()
构建网络
残差网络结构(Residual Network)是ResNet网络的主要亮点,ResNet使用残差网络结构后可有效地减轻退化问题,实现更深的网络结构设计,提高网络的训练精度。本节首先讲述如何构建残差网络结构,然后通过堆叠残差网络来构建ResNet50网络。
构建残差网络结构
残差网络结构图如下图所示,残差网络由两个分支构成:一个主分支,一个shortcuts(图中弧线表示)。主分支通过堆叠一系列的卷积操作得到,shotcuts从输入直接到输出,主分支输出的特征矩阵𝐹(𝑥)加上shortcuts输出的特征矩阵𝑥得到𝐹(𝑥)+𝑥,通过Relu激活函数后即为残差网络最后的输出。
残差网络结构主要由两种,一种是Building Block,适用于较浅的ResNet网络,如ResNet18和ResNet34;另一种是Bottleneck,适用于层数较深的ResNet网络,如ResNet50、ResNet101和ResNet152。
Building Block
Building Block结构图如下图所示,主分支有两层卷积网络结构:
- 主分支第一层网络以输入channel为64为例,首先通过一个3×3的卷积层,然后通过Batch Normalization层,最后通过Relu激活函数层,输出channel为64;
- 主分支第二层网络的输入channel为64,首先通过一个3×3的卷积层,然后通过Batch Normalization层,输出channel为64。
最后将主分支输出的特征矩阵与shortcuts输出的特征矩阵相加,通过Relu激活函数即为Building Block最后的输出。
主分支与shortcuts输出的特征矩阵相加时,需要保证主分支与shortcuts输出的特征矩阵shape相同。如果主分支与shortcuts输出的特征矩阵shape不相同,如输出channel是输入channel的一倍时,shortcuts上需要使用数量与输出channel相等,大小为1×1的卷积核进行卷积操作;若输出的图像较输入图像缩小一倍,则要设置shortcuts中卷积操作中的stride
为2,主分支第一层卷积操作的stride
也需设置为2。
from typing import Type, Union, List, Optional
import mindspore.nn as nn
from mindspore.common.initializer import Normal
# 初始化卷积层与BatchNorm的参数
weight_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 = norm
self.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_sample
def 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
Bottleneck
Bottleneck结构图如下图所示,在输入相同的情况下Bottleneck结构相对Building Block结构的参数数量更少,更适合层数较深的网络,ResNet50使用的残差结构就是Bottleneck。该结构的主分支有三层卷积结构,分别为1×1的卷积层、3×3卷积层和1×1的卷积层,其中1×1的卷积层分别起降维和升维的作用。
- 主分支第一层网络以输入channel为256为例,首先通过数量为64,大小为1×1的卷积核进行降维,然后通过Batch Normalization层,最后通过Relu激活函数层,其输出channel为64;
- 主分支第二层网络通过数量为64,大小为3×3的卷积核提取特征,然后通过Batch Normalization层,最后通过Relu激活函数层,其输出channel为64;
- 主分支第三层通过数量为256,大小1×1的卷积核进行升维,然后通过Batch Normalization层,其输出channel为256。
最后将主分支输出的特征矩阵与shortcuts输出的特征矩阵相加,通过Relu激活函数即为Bottleneck最后的输出。
主分支与shortcuts输出的特征矩阵相加时,需要保证主分支与shortcuts输出的特征矩阵shape相同。如果主分支与shortcuts输出的特征矩阵shape不相同,如输出channel是输入channel的一倍时,shortcuts上需要使用数量与输出channel相等,大小为1×1的卷积核进行卷积操作;若输出的图像较输入图像缩小一倍,则要设置shortcuts中卷积操作中的stride
为2,主分支第二层卷积操作的stride
也需设置为2。
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_sample
def 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
构建ResNet50网络
ResNet网络层结构如下图所示,以输入彩色图像224×224为例,首先通过数量64,卷积核大小为7×7,stride为2的卷积层conv1,该层输出图片大小为112×112,输出channel为64;然后通过一个3×3的最大下采样池化层,该层输出图片大小为56×56,输出channel为64;再堆叠4个残差网络块(conv2_x、conv3_x、conv4_x和conv5_x),此时输出图片大小为7×7,输出channel为2048;最后通过一个平均池化层、全连接层和softmax,得到分类概率。
对于每个残差网络块,以ResNet50网络中的conv2_x为例,其由3个Bottleneck结构堆叠而成,每个Bottleneck输入的channel为64,输出channel为256。
如下示例定义make_layer
实现残差块的构建,其参数如下所示:
last_out_channel
:上一个残差网络输出的通道数。block
:残差网络的类别,分别为ResidualBlockBase
和ResidualBlock
。channel
:残差网络输入的通道数。block_nums
:残差网络块堆叠的个数。stride
:卷积移动的步幅。
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)
ResNet50网络共有5个卷积结构,一个平均池化层,一个全连接层,以CIFAR-10数据集为例:
- conv1:输入图片大小为32×32,输入channel为3。首先经过一个卷积核数量为64,卷积核大小为7×7,stride为2的卷积层;然后通过一个Batch Normalization层;最后通过Reul激活函数。该层输出feature map大小为16×16,输出channel为64。
- conv2_x:输入feature map大小为16×16,输入channel为64。首先经过一个卷积核大小为3×3,stride为2的最大下采样池化操作;然后堆叠3个[1×1,64;3×3,64;1×1,256]结构的Bottleneck。该层输出feature map大小为8×88×8,输出channel为256。
- conv3_x:输入feature map大小为8×8,输入channel为256。该层堆叠4个[1×1,128;3×3,128;1×1,512]结构的Bottleneck。该层输出feature map大小为4×4,输出channel为512。
- conv4_x:输入feature map大小为4×4,输入channel为512。该层堆叠6个[1×1,256;3×3,256;1×1,1024]结构的Bottleneck。该层输出feature map大小为2×2,输出channel为1024。
- conv5_x:输入feature map大小为2×2,输入channel为1024。该层堆叠3个[1×1,512;3×3,512;1×1,2048]结构的Bottleneck。该层输出feature map大小为1×1,输出channel为2048。
- average pool & fc:输入channel为2048,输出channel为分类的类别数。
如下示例代码实现ResNet50模型的构建,通过用调函数resnet50
即可构建ResNet50模型,函数resnet50
参数如下:
num_classes
:分类的类别数,默认类别数为1000。pretrained
:下载对应的训练模型,并加载预训练模型中的参数到网络中。
from mindspore import load_checkpoint, load_param_into_net
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, pretrained_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=pretrained_ckpt, replace=True)
param_dict = load_checkpoint(pretrained_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)
模型训练与评估
本节使用ResNet50预训练模型进行微调。调用resnet50
构造ResNet50模型,并设置pretrained
参数为True,将会自动下载ResNet50预训练模型,并加载预训练模型中的参数到网络中。然后定义优化器和损失函数,逐个epoch打印训练的损失值和评估精度,并保存评估精度最高的ckpt文件(resnet50-best.ckpt)到当前路径的./BestCheckPoint下。
由于预训练模型全连接层(fc)的输出大小(对应参数num_classes
)为1000, 为了成功加载预训练权重,我们将模型的全连接输出大小设置为默认的1000。CIFAR10数据集共有10个分类,在使用该数据集进行训练时,需要将加载好预训练权重的模型全连接层输出大小重置为10。
# 定义ResNet50网络
network = resnet50(pretrained=True)
# 全连接层输入层的大小
in_channel = network.fc.in_channels
fc = nn.Dense(in_channels=in_channel, out_channels=10)
# 重置全连接层
network.fc = fc
# 设置学习率
num_epochs = 5
lr = nn.cosine_decay_lr(min_lr=0.00001, max_lr=0.001, total_step=step_size_train * num_epochs,
step_per_epoch=step_size_train, decay_epoch=num_epochs)
# 定义优化器和损失函数
opt = nn.Momentum(params=network.trainable_params(), learning_rate=lr, momentum=0.9)
loss_fn = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
def forward_fn(inputs, targets):
logits = network(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
import os
# 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)
# 最佳模型存储路径
best_acc = 0
best_ckpt_dir = "./BestCheckpoint"
best_ckpt_path = "./BestCheckpoint/resnet50-best.ckpt"
if not os.path.exists(best_ckpt_dir):
os.mkdir(best_ckpt_dir)
import mindspore.ops as ops
def train(data_loader, epoch):
"""模型训练"""
losses = []
network.set_train(True)
for i, (images, labels) in enumerate(data_loader):
loss = train_step(images, labels)
if i % 100 == 0 or i == step_size_train - 1:
print('Epoch: [%3d/%3d], Steps: [%3d/%3d], Train Loss: [%5.3f]' %
(epoch + 1, num_epochs, i + 1, step_size_train, loss))
losses.append(loss)
return sum(losses) / len(losses)
def evaluate(data_loader):
"""模型验证"""
network.set_train(False)
correct_num = 0.0 # 预测正确个数
total_num = 0.0 # 预测总数
for images, labels in data_loader:
logits = network(images)
pred = logits.argmax(axis=1) # 预测结果
correct = ops.equal(pred, labels).reshape((-1, ))
correct_num += correct.sum().asnumpy()
total_num += correct.shape[0]
acc = correct_num / total_num # 准确率
return acc
# 开始循环训练
print("Start Training Loop ...")
for epoch in range(num_epochs):
curr_loss = train(data_loader_train, epoch)
curr_acc = evaluate(data_loader_val)
print("-" * 50)
print("Epoch: [%3d/%3d], Average Train Loss: [%5.3f], Accuracy: [%5.3f]" % (
epoch+1, num_epochs, curr_loss, curr_acc
))
print("-" * 50)
# 保存当前预测准确率最高的模型
if curr_acc > best_acc:
best_acc = curr_acc
ms.save_checkpoint(network, 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)
可视化模型预测
定义visualize_model
函数,使用上述验证精度最高的模型对CIFAR-10测试数据集进行预测,并将预测结果可视化。若预测字体颜色为蓝色表示为预测正确,预测字体颜色为红色则表示预测错误。
import matplotlib.pyplot as plt
def visualize_model(best_ckpt_path, dataset_val):
num_class = 10 # 对狼和狗图像进行二分类
net = resnet50(num_class)
# 加载模型参数
param_dict = ms.load_checkpoint(best_ckpt_path)
ms.load_param_into_net(net, param_dict)
# 加载验证集的数据进行验证
data = next(dataset_val.create_dict_iterator())
images = data["image"]
labels = data["label"]
# 预测图像类别
output = net(data['image'])
pred = np.argmax(output.asnumpy(), axis=1)
# 图像分类
classes = []
with open(data_dir + "/batches.meta.txt", "r") as f:
for line in f:
line = line.rstrip()
if line:
classes.append(line)
# 显示图像及图像的预测值
plt.figure()
for i in range(6):
plt.subplot(2, 3, 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.asnumpy()[i] else 'red'
plt.title('predict:{}'.format(classes[pred[i]]), color=color)
picture_show = np.transpose(images.asnumpy()[i], (1, 2, 0))
mean = np.array([0.4914, 0.4822, 0.4465])
std = np.array([0.2023, 0.1994, 0.2010])
picture_show = std * picture_show + mean
picture_show = np.clip(picture_show, 0, 1)
plt.imshow(picture_show)
plt.axis('off')
plt.show()
# 使用测试数据集进行验证
visualize_model(best_ckpt_path=best_ckpt_path, dataset_val=dataset_val)
ShuffleNet网络介绍
ShuffleNetV1是旷视科技提出的一种计算高效的CNN模型,和MobileNet, SqueezeNet等一样主要应用在移动端,所以模型的设计目标就是利用有限的计算资源来达到最好的模型精度。ShuffleNetV1的设计核心是引入了两种操作:Pointwise Group Convolution和Channel Shuffle,这在保持精度的同时大大降低了模型的计算量。因此,ShuffleNetV1和MobileNet类似,都是通过设计更高效的网络结构来实现模型的压缩和加速。
了解ShuffleNet更多详细内容,详见论文ShuffleNet。
如下图所示,ShuffleNet在保持不低的准确率的前提下,将参数量几乎降低到了最小,因此其运算速度较快,单位参数量对模型准确率的贡献非常高。
模型架构
ShuffleNet最显著的特点在于对不同通道进行重排来解决Group Convolution带来的弊端。通过对ResNet的Bottleneck单元进行改进,在较小的计算量的情况下达到了较高的准确率。
Pointwise Group Convolution
Group Convolution(分组卷积)原理如下图所示,相比于普通的卷积操作,分组卷积的情况下,每一组的卷积核大小为in_channels/g*k*k,一共有g组,所有组共有(in_channels/g*k*k)*out_channels个参数,是正常卷积参数的1/g。分组卷积中,每个卷积核只处理输入特征图的一部分通道,其优点在于参数量会有所降低,但输出通道数仍等于卷积核的数量。
Depthwise Convolution(深度可分离卷积)将组数g分为和输入通道相等的in_channels
,然后对每一个in_channels
做卷积操作,每个卷积核只处理一个通道,记卷积核大小为1*k*k,则卷积核参数量为:in_channels*k*k,得到的feature maps通道数与输入通道数相等;
Pointwise Group Convolution(逐点分组卷积)在分组卷积的基础上,令每一组的卷积核大小为 1×1,卷积核参数量为(in_channels/g*1*1)*out_channels。
模型
ShuffleNet最显著的特点在于对不同通道进行重排来解决Group Convolution带来的弊端。通过对ResNet的Bottleneck单元进行改进,在较小的计算量的情况下达到了较高的准确率。
Pointwise Group Convolution
Group Convolution(分组卷积)原理如下图所示,相比于普通的卷积操作,分组卷积的情况下,每一组的卷积核大小为in_channels/g*k*k,一共有g组,所有组共有(in_channels/g*k*k)*out_channels个参数,是正常卷积参数的1/g。分组卷积中,每个卷积核只处理输入特征图的一部分通道,其优点在于参数量会有所降低,但输出通道数仍等于卷积核的数量。
图片来源:Huang G, Liu S, Van der Maaten L, et al. Condensenet: An efficient densenet using learned group convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 2752-2761.
Depthwise Convolution(深度可分离卷积)将组数g分为和输入通道相等的in_channels
,然后对每一个in_channels
做卷积操作,每个卷积核只处理一个通道,记卷积核大小为1*k*k,则卷积核参数量为:in_channels*k*k,得到的feature maps通道数与输入通道数相等;
Pointwise Group Convolution(逐点分组卷积)在分组卷积的基础上,令每一组的卷积核大小为 1×1,卷积核参数量为(in_channels/g*1*1)*out_channels。
from mindspore import nn
import mindspore.ops as ops
from mindspore import Tensor
class GroupConv(nn.Cell):
def __init__(self, in_channels, out_channels, kernel_size,
stride, pad_mode="pad", pad=0, groups=1, has_bias=False):
super(GroupConv, self).__init__()
self.groups = groups
self.convs = nn.CellList()
for _ in range(groups):
self.convs.append(nn.Conv2d(in_channels // groups, out_channels // groups,
kernel_size=kernel_size, stride=stride, has_bias=has_bias,
padding=pad, pad_mode=pad_mode, group=1, weight_init='xavier_uniform'))
def construct(self, x):
features = ops.split(x, split_size_or_sections=int(len(x[0]) // self.groups), axis=1)
outputs = ()
for i in range(self.groups):
outputs = outputs + (self.convs[i](features[i].astype("float32")),)
out = ops.cat(outputs, axis=1)
return out
Channel Shuffle
Group Convolution的弊端在于不同组别的通道无法进行信息交流,堆积GConv层后一个问题是不同组之间的特征图是不通信的,这就好像分成了g个互不相干的道路,每一个人各走各的,这可能会降低网络的特征提取能力。这也是Xception,MobileNet等网络采用密集的1x1卷积(Dense Pointwise Convolution)的原因。
为了解决不同组别通道“近亲繁殖”的问题,ShuffleNet优化了大量密集的1x1卷积(在使用的情况下计算量占用率达到了惊人的93.4%),引入Channel Shuffle机制(通道重排)。这项操作直观上表现为将不同分组通道均匀分散重组,使网络在下一层能处理不同组别通道的信息。
如下图所示,对于g组,每组有n个通道的特征图,首先reshape成g行n列的矩阵,再将矩阵转置成n行g列,最后进行flatten操作,得到新的排列。这些操作都是可微分可导的且计算简单,在解决了信息交互的同时符合了ShuffleNet轻量级网络设计的轻量特征。
为了阅读方便,将Channel Shuffle的代码实现放在下方ShuffleNet模块的代码中。
ShuffleNet模块
如下图所示,ShuffleNet对ResNet中的Bottleneck结构进行由(a)到(b), (c)的更改:
-
将开始和最后的1×1卷积模块(降维、升维)改成Point Wise Group Convolution;
-
为了进行不同通道的信息交流,再降维之后进行Channel Shuffle;
-
降采样模块中,3×3 Depth Wise Convolution的步长设置为2,长宽降为原来的一般,因此shortcut中采用步长为2的3×3平均池化,并把相加改成拼接。
class ShuffleV1Block(nn.Cell):
def __init__(self, inp, oup, group, first_group, mid_channels, ksize, stride):
super(ShuffleV1Block, self).__init__()
self.stride = stride
pad = ksize // 2
self.group = group
if stride == 2:
outputs = oup - inp
else:
outputs = oup
self.relu = nn.ReLU()
branch_main_1 = [
GroupConv(in_channels=inp, out_channels=mid_channels,
kernel_size=1, stride=1, pad_mode="pad", pad=0,
groups=1 if first_group else group),
nn.BatchNorm2d(mid_channels),
nn.ReLU(),
]
branch_main_2 = [
nn.Conv2d(mid_channels, mid_channels, kernel_size=ksize, stride=stride,
pad_mode='pad', padding=pad, group=mid_channels,
weight_init='xavier_uniform', has_bias=False),
nn.BatchNorm2d(mid_channels),
GroupConv(in_channels=mid_channels, out_channels=outputs,
kernel_size=1, stride=1, pad_mode="pad", pad=0,
groups=group),
nn.BatchNorm2d(outputs),
]
self.branch_main_1 = nn.SequentialCell(branch_main_1)
self.branch_main_2 = nn.SequentialCell(branch_main_2)
if stride == 2:
self.branch_proj = nn.AvgPool2d(kernel_size=3, stride=2, pad_mode='same')
def construct(self, old_x):
left = old_x
right = old_x
out = old_x
right = self.branch_main_1(right)
if self.group > 1:
right = self.channel_shuffle(right)
right = self.branch_main_2(right)
if self.stride == 1:
out = self.relu(left + right)
elif self.stride == 2:
left = self.branch_proj(left)
out = ops.cat((left, right), 1)
out = self.relu(out)
return out
def channel_shuffle(self, x):
batchsize, num_channels, height, width = ops.shape(x)
group_channels = num_channels // self.group
x = ops.reshape(x, (batchsize, group_channels, self.group, height, width))
x = ops.transpose(x, (0, 2, 1, 3, 4))
x = ops.reshape(x, (batchsize, num_channels, height, width))
return x
构建ShuffleNet网络
ShuffleNet网络结构如下图所示,以输入图像224×224,组数3(g = 3)为例,首先通过数量24,卷积核大小为3×3,stride为2的卷积层,输出特征图大小为112×112,channel为24;然后通过stride为2的最大池化层,输出特征图大小为56×56,channel数不变;再堆叠3个ShuffleNet模块(Stage2, Stage3, Stage4),三个模块分别重复4次、8次、4次,其中每个模块开始先经过一次下采样模块(上图(c)),使特征图长宽减半,channel翻倍(Stage2的下采样模块除外,将channel数从24变为240);随后经过全局平均池化,输出大小为1×1×960,再经过全连接层和softmax,得到分类概率。
class ShuffleNetV1(nn.Cell):
def __init__(self, n_class=1000, model_size='2.0x', group=3):
super(ShuffleNetV1, self).__init__()
print('model size is ', model_size)
self.stage_repeats = [4, 8, 4]
self.model_size = model_size
if group == 3:
if model_size == '0.5x':
self.stage_out_channels = [-1, 12, 120, 240, 480]
elif model_size == '1.0x':
self.stage_out_channels = [-1, 24, 240, 480, 960]
elif model_size == '1.5x':
self.stage_out_channels = [-1, 24, 360, 720, 1440]
elif model_size == '2.0x':
self.stage_out_channels = [-1, 48, 480, 960, 1920]
else:
raise NotImplementedError
elif group == 8:
if model_size == '0.5x':
self.stage_out_channels = [-1, 16, 192, 384, 768]
elif model_size == '1.0x':
self.stage_out_channels = [-1, 24, 384, 768, 1536]
elif model_size == '1.5x':
self.stage_out_channels = [-1, 24, 576, 1152, 2304]
elif model_size == '2.0x':
self.stage_out_channels = [-1, 48, 768, 1536, 3072]
else:
raise NotImplementedError
input_channel = self.stage_out_channels[1]
self.first_conv = nn.SequentialCell(
nn.Conv2d(3, input_channel, 3, 2, 'pad', 1, weight_init='xavier_uniform', has_bias=False),
nn.BatchNorm2d(input_channel),
nn.ReLU(),
)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode='same')
features = []
for idxstage in range(len(self.stage_repeats)):
numrepeat = self.stage_repeats[idxstage]
output_channel = self.stage_out_channels[idxstage + 2]
for i in range(numrepeat):
stride = 2 if i == 0 else 1
first_group = idxstage == 0 and i == 0
features.append(ShuffleV1Block(input_channel, output_channel,
group=group, first_group=first_group,
mid_channels=output_channel // 4, ksize=3, stride=stride))
input_channel = output_channel
self.features = nn.SequentialCell(features)
self.globalpool = nn.AvgPool2d(7)
self.classifier = nn.Dense(self.stage_out_channels[-1], n_class)
def construct(self, x):
x = self.first_conv(x)
x = self.maxpool(x)
x = self.features(x)
x = self.globalpool(x)
x = ops.reshape(x, (-1, self.stage_out_channels[-1]))
x = self.classifier(x)
return x
模型训练和评估
采用CIFAR-10数据集对ShuffleNet进行预训练。
训练集准备与加载
采用CIFAR-10数据集对ShuffleNet进行预训练。CIFAR-10共有60000张32*32的彩色图像,均匀地分为10个类别,其中50000张图片作为训练集,10000图片作为测试集
from download import download
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/cifar-10-binary.tar.gz"
download(url, "./dataset", kind="tar.gz", replace=True)
import mindspore as ms
from mindspore.dataset import Cifar10Dataset
from mindspore.dataset import vision, transforms
def get_dataset(train_dataset_path, batch_size, usage):
image_trans = []
if usage == "train":
image_trans = [
vision.RandomCrop((32, 32), (4, 4, 4, 4)),
vision.RandomHorizontalFlip(prob=0.5),
vision.Resize((224, 224)),
vision.Rescale(1.0 / 255.0, 0.0),
vision.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]),
vision.HWC2CHW()
]
elif usage == "test":
image_trans = [
vision.Resize((224, 224)),
vision.Rescale(1.0 / 255.0, 0.0),
vision.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]),
vision.HWC2CHW()
]
label_trans = transforms.TypeCast(ms.int32)
dataset = Cifar10Dataset(train_dataset_path, usage=usage, shuffle=True)
dataset = dataset.map(image_trans, 'image')
dataset = dataset.map(label_trans, 'label')
dataset = dataset.batch(batch_size, drop_remainder=True)
return dataset
dataset = get_dataset("./dataset/cifar-10-batches-bin", 128, "train")
batches_per_epoch = dataset.get_dataset_size()
模型训练
本节用随机初始化的参数做预训练。首先调用ShuffleNetV1
定义网络,参数量选择"2.0x"
,并定义损失函数为交叉熵损失,学习率经过4轮的warmup
后采用余弦退火,优化器采用Momentum
。最后用train.model
中的Model
接口将模型、损失函数、优化器封装在model
中,并用model.train()
对网络进行训练。将ModelCheckpoint
、CheckpointConfig
、TimeMonitor
和LossMonitor
传入回调函数中,将会打印训练的轮数、损失和时间,并将ckpt文件保存在当前目录下。
import time
import mindspore
import numpy as np
from mindspore import Tensor, nn
from mindspore.train import ModelCheckpoint, CheckpointConfig, TimeMonitor, LossMonitor, Model, Top1CategoricalAccuracy, Top5CategoricalAccuracy
def train():
mindspore.set_context(mode=mindspore.PYNATIVE_MODE, device_target="Ascend")
net = ShuffleNetV1(model_size="2.0x", n_class=10)
loss = nn.CrossEntropyLoss(weight=None, reduction='mean', label_smoothing=0.1)
min_lr = 0.0005
base_lr = 0.05
lr_scheduler = mindspore.nn.cosine_decay_lr(min_lr,
base_lr,
batches_per_epoch*250,
batches_per_epoch,
decay_epoch=250)
lr = Tensor(lr_scheduler[-1])
optimizer = nn.Momentum(params=net.trainable_params(), learning_rate=lr, momentum=0.9, weight_decay=0.00004, loss_scale=1024)
loss_scale_manager = ms.amp.FixedLossScaleManager(1024, drop_overflow_update=False)
model = Model(net, loss_fn=loss, optimizer=optimizer, amp_level="O3", loss_scale_manager=loss_scale_manager)
callback = [TimeMonitor(), LossMonitor()]
save_ckpt_path = "./"
config_ckpt = CheckpointConfig(save_checkpoint_steps=batches_per_epoch, keep_checkpoint_max=5)
ckpt_callback = ModelCheckpoint("shufflenetv1", directory=save_ckpt_path, config=config_ckpt)
callback += [ckpt_callback]
print("============== Starting Training ==============")
start_time = time.time()
# 由于时间原因,epoch = 5,可根据需求进行调整
model.train(5, dataset, callbacks=callback)
use_time = time.time() - start_time
hour = str(int(use_time // 60 // 60))
minute = str(int(use_time // 60 % 60))
second = str(int(use_time % 60))
print("total time:" + hour + "h " + minute + "m " + second + "s")
print("============== Train Success ==============")
if __name__ == '__main__':
train()
模型评估
在CIFAR-10的测试集上对模型进行评估。
设置好评估模型的路径后加载数据集,并设置Top 1, Top 5的评估标准,最后用model.eval()
接口对模型进行评估。
from mindspore import load_checkpoint, load_param_into_net
def test():
mindspore.set_context(mode=mindspore.GRAPH_MODE, device_target="Ascend")
dataset = get_dataset("./dataset/cifar-10-batches-bin", 128, "test")
net = ShuffleNetV1(model_size="2.0x", n_class=10)
param_dict = load_checkpoint("shufflenetv1-5_390.ckpt")
load_param_into_net(net, param_dict)
net.set_train(False)
loss = nn.CrossEntropyLoss(weight=None, reduction='mean', label_smoothing=0.1)
eval_metrics = {'Loss': nn.Loss(), 'Top_1_Acc': Top1CategoricalAccuracy(),
'Top_5_Acc': Top5CategoricalAccuracy()}
model = Model(net, loss_fn=loss, metrics=eval_metrics)
start_time = time.time()
res = model.eval(dataset, dataset_sink_mode=False)
use_time = time.time() - start_time
hour = str(int(use_time // 60 // 60))
minute = str(int(use_time // 60 % 60))
second = str(int(use_time % 60))
log = "result:" + str(res) + ", ckpt:'" + "./shufflenetv1-5_390.ckpt" \
+ "', time: " + hour + "h " + minute + "m " + second + "s"
print(log)
filename = './eval_log.txt'
with open(filename, 'a') as file_object:
file_object.write(log + '\n')
if __name__ == '__main__':
test()
模型预测
在CIFAR-10的测试集上对模型进行预测,并将预测结果可视化。
import mindspore
import matplotlib.pyplot as plt
import mindspore.dataset as ds
net = ShuffleNetV1(model_size="2.0x", n_class=10)
show_lst = []
param_dict = load_checkpoint("shufflenetv1-5_390.ckpt")
load_param_into_net(net, param_dict)
model = Model(net)
dataset_predict = ds.Cifar10Dataset(dataset_dir="./dataset/cifar-10-batches-bin", shuffle=False, usage="train")
dataset_show = ds.Cifar10Dataset(dataset_dir="./dataset/cifar-10-batches-bin", shuffle=False, usage="train")
dataset_show = dataset_show.batch(16)
show_images_lst = next(dataset_show.create_dict_iterator())["image"].asnumpy()
image_trans = [
vision.RandomCrop((32, 32), (4, 4, 4, 4)),
vision.RandomHorizontalFlip(prob=0.5),
vision.Resize((224, 224)),
vision.Rescale(1.0 / 255.0, 0.0),
vision.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]),
vision.HWC2CHW()
]
dataset_predict = dataset_predict.map(image_trans, 'image')
dataset_predict = dataset_predict.batch(16)
class_dict = {0:"airplane", 1:"automobile", 2:"bird", 3:"cat", 4:"deer", 5:"dog", 6:"frog", 7:"horse", 8:"ship", 9:"truck"}
# 推理效果展示(上方为预测的结果,下方为推理效果图片)
plt.figure(figsize=(16, 5))
predict_data = next(dataset_predict.create_dict_iterator())
output = model.predict(ms.Tensor(predict_data['image']))
pred = np.argmax(output.asnumpy(), axis=1)
index = 0
for image in show_images_lst:
plt.subplot(2, 8, index+1)
plt.title('{}'.format(class_dict[pred[index]]))
index += 1
plt.imshow(image)
plt.axis("off")
plt.show()
最后打卡今天的学习时间
心得
今天是昇思25天学习打卡营的第十一天,我学习了ResNet50和ShuffleNet的图像分类方法。ResNet50通过引入残差网络结构有效解决了深层网络的退化问题,使得网络层数可以突破1000层,并在CIFAR-10数据集上表现出色。而ShuffleNetV1则是一种专为移动端设计的计算高效的CNN模型,通过引入Pointwise Group Convolution和Channel Shuffle,大大降低了计算量,同时保持了模型的高精度。通过对这两种网络架构的学习,我深入理解了深度学习中不同的网络优化方法及其在实际应用中的优势。