昇思25天学习打卡营第12天|ShuffleNet图像分类

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。

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。

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 $1\times 1$，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), ©的更改：

1. 将开始和最后的$1\times 1$卷积模块（降维、升维）改成Point Wise Group Convolution；

2. 为了进行不同通道的信息交流，再降维之后进行Channel Shuffle；

3. 降采样模块中，$3\times 3$ 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$3\times 3$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$224\times 224$，组数3（g = 3）为例，首先通过数量24，卷积核大小为$3\times 3$，stride为2的卷积层，输出特征图大小为$112\times 112$，channel为24；然后通过stride为2的最大池化层，输出特征图大小为$56\times 56$，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©），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$1\times 1\times 960$，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

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()