华为开源自研AI框架昇思MindSpore应用案例：计算高效的卷积模型ShuffleNet

如果你对MindSpore感兴趣，可以关注昇思MindSpore社区

ShuffleNet

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：Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, Jian Sun. "ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.

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。

图片来源：Bianco S, Cadene R, Celona L, et al. Benchmark analysis of representative deep neural network architectures[J]. IEEE access, 2018, 6: 64270-64277.

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1×1，卷积核参数量为(in_channels/g*1*1)*out_channels。

In [1]:

from mindspore import nn
import mindspore.ops.operations as P
from mindspore import dtype as mstypeclass 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 = groupsself.convs = nn.CellList()self.op_split = P.Split(axis=1, output_num=self.groups)     # 分割成groups组self.op_concat = P.Concat(axis=1)self.cast = P.Cast()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 = self.op_split(x)     # 将输入x按通道拆分成groups组outputs = ()for i in range(self.groups):outputs = outputs + (self.convs[i](self.cast(features[i], mstype.float32)),)out = self.op_concat(outputs)   # 最后拼接起来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：

1. 将开始和最后的1×1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，3×3 depthwise 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。

In [2]:

class ShuffleV1Block(nn.Cell):def __init__(self, inp, oup, group, first_group, mid_channels, ksize, stride):super(ShuffleV1Block, self).__init__()self.stride = stridepad = ksize // 2self.group = groupif stride == 2:outputs = oup - inpelse:outputs = oupself.relu = nn.ReLU()self.add = P.Add()self.concat = P.Concat(1)self.shape = P.Shape()self.transpose = P.Transpose()self.reshape = P.Reshape()branch_main_1 = [# pointwise group convolutionGroupConv(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 = [# depthwise group convolutionnn.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),# pointwise group convolutionGroupConv(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_xright = old_xout = old_xright = 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(self.add(left, right))elif self.stride == 2:left = self.branch_proj(left)out = self.concat((left, right))out = self.relu(out)return outdef channel_shuffle(self, x):batchsize, num_channels, height, width = self.shape(x)group_channels = num_channels // self.groupx = self.reshape(x, (batchsize, group_channels, self.group, height, width))x = self.transpose(x, (0, 2, 1, 3, 4))x = self.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。

In [3]:

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_sizeif 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 NotImplementedErrorelif 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# building first layerinput_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 1first_group = idxstage == 0 and i == 0features.append(ShuffleV1Block(input_channel, output_channel,group=group, first_group=first_group,mid_channels=output_channel // 4, ksize=3, stride=stride))input_channel = output_channelself.features = nn.SequentialCell(features)self.globalpool = nn.AvgPool2d(7)self.classifier = nn.Dense(self.stage_out_channels[-1], n_class)self.reshape = P.Reshape()def construct(self, x):x = self.first_conv(x)x = self.maxpool(x)x = self.features(x)x = self.globalpool(x)x = self.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。并且为了测试模型在样本数量较少，但尺寸较大的数据集上的迁移性能，用flower_photos数据集对模型进行微调。

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。如下示例使用mindspore.dataset.Cifar10Dataset接口下载并加载CIFAR-10的训练集。目前仅支持二进制版本（CIFAR-10 binary version）。

In [4]:

import mindspore as ms
from mindspore.dataset import Cifar10Dataset
from mindspore.dataset import vision, transformsdef get_dataset(train_dataset_path, batch_size, usage):image_trans = []if usage=="train":image_trans = [vision.c_transforms.RandomCrop((32, 32), (4, 4, 4, 4)),vision.c_transforms.RandomHorizontalFlip(prob=0.5),vision.c_transforms.Resize((224, 224)),vision.c_transforms.Rescale(1.0 / 255.0, 0.0),vision.c_transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]),vision.c_transforms.HWC2CHW()]elif usage=="test":image_trans = [vision.c_transforms.Resize((224, 224)),vision.c_transforms.Rescale(1.0 / 255.0, 0.0),vision.c_transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]),vision.c_transforms.HWC2CHW()]label_trans = transforms.c_transforms.TypeCast(ms.int32)dataset = Cifar10Dataset(train_dataset_path, usage=usage, shuffle=True if usage == "train" else False)dataset = dataset.map(image_trans, 'image')dataset = dataset.map(label_trans, 'label')dataset = dataset.batch(batch_size)return dataset

In [5]:

dataset = get_dataset("./data/cifar10/cifar-10-batches-bin/", 128, "train")
batches_per_epoch = dataset.get_dataset_size()

数据集文件目录结构如下：

In [ ]:

./datasets
└── cifar-10-batches-bin├── batches.meta.txt├── data_batch_1.bin├── data_batch_2.bin├── data_batch_3.bin├── data_batch_4.bin├── data_batch_5.bin├── readme.html└── test_batch.bin

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, 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。

In [ ]:

import os
import time
import math
import mindspore
import numpy as np
from mindspore import Tensor, nn
from mindspore.nn.optim.momentum import Momentum
from mindspore.train.model import Model
from src.crossentropysmooth import CrossEntropySmooth
from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, TimeMonitor, LossMonitor
from mindspore.train.loss_scale_manager import FixedLossScaleManagerdef get_cosine_lr(lr_init, lr_end, lr_max, warmup_epochs, total_epochs, steps_per_epoch):total_steps = steps_per_epoch * total_epochswarmup_steps = steps_per_epoch * warmup_epochsdecay_steps = total_steps - warmup_stepslr_each_step = []for i in range(total_steps):if i < warmup_steps:lr_inc = (float(lr_max) - float(lr_init)) / float(warmup_steps)lr = float(lr_init) + lr_inc * (i + 1)else:cosine_decay = 0.5 * (1 + math.cos(math.pi * (i-warmup_steps) / decay_steps))lr = (lr_max-lr_end)*cosine_decay + lr_endlr_each_step.append(lr)return lr_each_stepdef train():# Ascend训练，若用GPU，请改成"GPU"device_target = "GPU"mindspore.set_context(mode=mindspore.GRAPH_MODE, device_target=device_target, save_graphs=False)if device_target == "GPU":mindspore.set_context(enable_graph_kernel=True)# context.set_context(device_id=0)# context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False)# define networknet = ShuffleNetV1(model_size="2.0x", n_class=10)# define lossloss = CrossEntropySmooth(sparse=True, reduction="mean", smooth_factor=0.1, num_classes=10)# get learning ratelr = get_cosine_lr(lr_init=0.00, lr_end=0.50, lr_max=0.50, warmup_epochs=4,total_epochs=250, steps_per_epoch=batches_per_epoch)lr = Tensor(lr)# define optimizationoptimizer = Momentum(params=net.trainable_params(), learning_rate=lr, momentum=0.9,weight_decay=0.00004, loss_scale=1024)# modelloss_scale_manager = FixedLossScaleManager(1024, drop_overflow_update=False)model = Model(net, loss_fn=loss, optimizer=optimizer, amp_level="O3",loss_scale_manager=loss_scale_manager)# define callbackscb = [TimeMonitor(), LossMonitor()]save_ckpt_path = "./"config_ck = CheckpointConfig(save_checkpoint_steps=batches_per_epoch, keep_checkpoint_max=5)ckpt_cb = ModelCheckpoint("shufflenetv1", directory=save_ckpt_path, config=config_ck)cb += [ckpt_cb]print("============== Starting Training ==============")start_time = time.time()# begin trainmodel.train(250, dataset, callbacks=cb, dataset_sink_mode=True)use_time = time.time() - start_timehour = 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()

In [ ]:

============== Starting Training ==============
epoch: 1 step: 195, loss is 1.7743151187896729
Train epoch time: 191255.396 ms, per step time: 980.797 ms
epoch: 2 step: 195, loss is 1.434311866760254
Train epoch time: 85797.786 ms, per step time: 439.989 ms
epoch: 3 step: 195, loss is 1.3389689922332764
Train epoch time: 83486.459 ms, per step time: 428.136 ms
epoch: 4 step: 195, loss is 1.2366113662719727
Train epoch time: 83831.257 ms, per step time: 429.904 ms
epoch: 5 step: 195, loss is 1.1475858688354492
Train epoch time: 84411.272 ms, per step time: 432.878 ms......epoch: 246 step: 195, loss is 0.5449619293212891
Train epoch time: 83016.999 ms, per step time: 425.728 ms
epoch: 247 step: 195, loss is 0.5449709892272949
Train epoch time: 84569.001 ms, per step time: 433.687 ms
epoch: 248 step: 195, loss is 0.5449811220169067
Train epoch time: 82748.939 ms, per step time: 424.354 ms
epoch: 249 step: 195, loss is 0.5449773073196411
Train epoch time: 84571.723 ms, per step time: 433.701 ms
epoch: 250 step: 195, loss is 0.5450158715248108
Train epoch time: 84446.256 ms, per step time: 433.058 ms
total time:5h 50m 50s
============== Train Success ==============

训练好的模型保存在当前目录的shufflenetv1-250_195.ckpt中，用作评估。

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。

In [8]:

import time
import mindsporefrom mindspore import nn, set_context
from mindspore import load_checkpoint, load_param_into_net
from src.crossentropysmooth import CrossEntropySmooth
from mindspore.train.model import Modeldef test():set_context(mode=mindspore.GRAPH_MODE, device_target="Ascend")# 加载测试集dataset = get_dataset("./data/cifar10/cifar-10-batches-bin/", 128, "test")# define netnet = ShuffleNetV1(model_size="2.0x", n_class=10)# load checkpointparam_dict = load_checkpoint("./ckpt/shufflenetv1-250_195.ckpt")load_param_into_net(net, param_dict)net.set_train(False)# define lossloss = CrossEntropySmooth(sparse=True, reduction="mean", smooth_factor=0.1, num_classes=10)# define modeleval_metrics = {'Loss': nn.Loss(), 'Top_1_Acc': nn.Top1CategoricalAccuracy(),'Top_5_Acc': nn.Top5CategoricalAccuracy()}model = Model(net, loss_fn=loss, metrics=eval_metrics)# start evaluatingstart_time = time.time()res = model.eval(dataset, dataset_sink_mode=False)use_time = time.time() - start_timehour = str(int(use_time // 60 // 60))minute = str(int(use_time // 60 % 60))second = str(int(use_time % 60))log = "result:" + str(res) + ", ckpt:'" + "./shufflenetv1-250_195.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()

model size is  2.0x
result:{'Loss': 0.7009380474875245, 'Top_1_Acc': 0.9434, 'Top_5_Acc': 0.9933}, ckpt:'./shufflenetv1-250_195.ckpt', time: 0h 2m 46s

In [ ]:

result:{'Loss': 0.6998715905042795, 'Top_1_Acc': 0.9434094551282052, 'Top_5_Acc': 0.9932892628205128}, ckpt:'./shufflenetv1-250_195.ckpt', time: 0h 1m 27s

迁移训练

采用flower_photos数据集对预训练好的模型进行迁移训练。flower-photos共有3670张大小不一的彩色图片，分成数目不等的5个类别（每个类别图片数量分别为633, 898, 641, 699, 799张）。因此首先需要对图片进行预处理，将其中80%用于训练，20%用于验证。

In [ ]:

import mindspore.common.dtype as mstype
import mindspore.dataset as ds
import mindspore.dataset.transforms.c_transforms as C2
import mindspore.dataset.vision.c_transforms as Cdef create_flower_dataset():        # 迁移训练的数据集ds.config.set_seed(1)dataset = ds.ImageFolderDataset("./data/flower_photos", num_parallel_workers=6, shuffle=False)train_dataset, eval_dataset = dataset.split([0.8, 0.2], randomize=True)trans = [C.RandomCropDecodeResize(224),C.RandomHorizontalFlip(prob=0.5),C.RandomColorAdjust(brightness=0.4, contrast=0.4, saturation=0.4),C.Normalize(mean=[0.485 * 255, 0.456 * 255, 0.406 * 255], std=[0.229 * 255, 0.224 * 255, 0.225 * 255]),C.HWC2CHW()]evals = [C.Decode(),C.Resize(239),C.CenterCrop(224),C.Normalize(mean=[0.485 * 255, 0.456 * 255, 0.406 * 255], std=[0.229 * 255, 0.224 * 255, 0.225 * 255]),C.HWC2CHW()]type_cast_op = C2.TypeCast(mstype.int32)train_dataset = train_dataset.map(input_columns="image", operations=trans, num_parallel_workers=6)train_dataset = train_dataset.map(input_columns="label", operations=type_cast_op, num_parallel_workers=6)eval_dataset = eval_dataset.map(input_columns="image", operations=evals, num_parallel_workers=6)eval_dataset = eval_dataset.map(input_columns="label", operations=type_cast_op, num_parallel_workers=6)# apply batch operationstrain_dataset = train_dataset.batch(128, drop_remainder=True)eval_dataset = eval_dataset.batch(128, drop_remainder=True)return train_dataset, eval_dataset

数据集文件目录结构如下：

In [ ]:

./flower_photos/
├── daisy
├── dandelion
├── roses
├── sunflowers
└── tulips

迁移训练的实现和预训练几乎完全相同，区别是在batch_size=128的条件下训练100轮。 训练集上最高Top 1准确率达到了0.905，并将模型保存到./ckpt/shufflenetv1_transfer_best.ckpt中。

测试集上Top 1准确率也达到了0.904。

In [ ]:

result:{'Loss': 0.6745888233184815, 'Top_1_Acc': 0.9046875}, ckpt:'./ckpt/shufflenetv1_transfer_best.ckpt', time: 0h 1m 47s