自学内容网 自学内容网

昇思25天学习打卡营第23天|基于MobileNetV2的垃圾分类

昇思MindSpore应用实践

本系列文章主要用于记录昇思25天学习打卡营的学习心得。

1、MobileNetv2模型原理介绍

MobileNet网络是由Google团队于2017年提出的专注于移动端、嵌入式或IoT设备的轻量级CNN网络,相比于传统的卷积神经网络,MobileNet网络使用深度可分离卷积(Depthwise Separable Convolution)的思想在准确率小幅度降低的前提下,大大减小了模型参数与运算量。并引入宽度系数 α和分辨率系数 β使模型满足不同应用场景的需求。

由于MobileNet网络中Relu激活函数处理低维特征信息时会存在大量的丢失,所以MobileNetV2网络提出使用倒残差结构(Inverted residual block)和Linear Bottlenecks来设计网络,以提高模型的准确率,且优化后的模型更小。
在这里插入图片描述

图中Inverted residual block结构是先使用1x1卷积进行升维,然后使用3x3的DepthWise卷积,最后使用1x1的卷积进行降维,与Residual block结构相反。Residual block是先使用1x1的卷积进行降维,然后使用3x3的卷积,最后使用1x1的卷积进行升维。

除了图像分类,MobileNetV2也可以用于目标检测(将分类结果的空间信息用检测框显示)、图像语义分割任务(相当于像素级的图像分类任务):
在这里插入图片描述
改动思路如下:

特征提取的输出

1、在图像分类任务中,MobileNetV2的最后一层是全局平均池化层和1x1卷积,将特征图转换为固定长度的特征向量,用于分类。然而,目标检测和语义分割任务需要保留更多的空间信息和位置信息,需要改变特征提取器的输出。
2、对于目标检测,通常会在特征提取器后添加额外的卷积层或者特定的ROI池化层,以保留特征图的空间信息,使得检测算法可以在图像中定位和识别目标。
3、对于图像语义分割,需要对特征提取器的输出进行上采样操作,以便将低分辨率的特征图映射到原始图像的大小,从而生成像素级的语义分割结果。

输出层的调整

1、在图像分类任务中,输出层通常是一个全连接层,用于将特征向量映射到类别概率分布上。
2、对于目标检测任务,输出层需要改为适合检测任务的结构,例如多个边界框的回归和分类,可能会包括卷积层和后续的处理层来生成和调整检测框,可参考Yolo与SSD目标检测的输出层部分。
3、对于图像语义分割任务,输出层需要调整为生成每个像素的类别标签或者像素级别的类别预测,通常使用转置卷积操作来逐像素地预测。

损失函数的设计

1、图像分类任务通常使用交叉熵损失来衡量预测类别与真实标签之间的差异。
2、目标检测任务需要使用特定的损失函数,例如目标检测中常用的分类损失和边界框回归损失的组合,以及可能的区域建议网络(如Faster R-CNN中的RPN)所需的额外损失。
3、图像语义分割任务通常使用像素级别的损失函数,如交叉熵损失或者Dice损失,以确保生成的分割结果与真实分割图像尽可能接近。

详细内容可参见MobileNetV2论文

2、数据准备

MobileNetV2的代码默认使用ImageFolder格式管理数据集,每一类图片整理成单独的一个文件夹, 数据集结构如下:

└─ImageFolder

├─train
│   class1Folder
│   ......
└─eval
    class1Folder
    ......
import math
import numpy as np
import os
import random

from matplotlib import pyplot as plt
from easydict import EasyDict
from PIL import Image
import numpy as np
import mindspore.nn as nn
from mindspore import ops as P
from mindspore.ops import add
from mindspore import Tensor
import mindspore.common.dtype as mstype
import mindspore.dataset as de
import mindspore.dataset.vision as C
import mindspore.dataset.transforms as C2
import mindspore as ms
from mindspore import set_context, nn, Tensor, load_checkpoint, save_checkpoint, export
from mindspore.train import Model
from mindspore.train import Callback, LossMonitor, ModelCheckpoint, CheckpointConfig

os.environ['GLOG_v'] = '3' # Log level includes 3(ERROR), 2(WARNING), 1(INFO), 0(DEBUG).
os.environ['GLOG_logtostderr'] = '0' # 0:输出到文件,1:输出到屏幕
os.environ['GLOG_log_dir'] = '../../log' # 日志目录
os.environ['GLOG_stderrthreshold'] = '2' # 输出到目录也输出到屏幕:3(ERROR), 2(WARNING), 1(INFO), 0(DEBUG).
set_context(mode=ms.GRAPH_MODE, device_target="CPU", device_id=0) # 设置采用图模式执行,设备为Ascend#

# 垃圾分类数据集标签,以及用于标签映射的字典。
garbage_classes = {
    '干垃圾': ['贝壳', '打火机', '旧镜子', '扫把', '陶瓷碗', '牙刷', '一次性筷子', '脏污衣服'],
    '可回收物': ['报纸', '玻璃制品', '篮球', '塑料瓶', '硬纸板', '玻璃瓶', '金属制品', '帽子', '易拉罐', '纸张'],
    '湿垃圾': ['菜叶', '橙皮', '蛋壳', '香蕉皮'],
    '有害垃圾': ['电池', '药片胶囊', '荧光灯', '油漆桶']
}

class_cn = ['贝壳', '打火机', '旧镜子', '扫把', '陶瓷碗', '牙刷', '一次性筷子', '脏污衣服',
            '报纸', '玻璃制品', '篮球', '塑料瓶', '硬纸板', '玻璃瓶', '金属制品', '帽子', '易拉罐', '纸张',
            '菜叶', '橙皮', '蛋壳', '香蕉皮',
            '电池', '药片胶囊', '荧光灯', '油漆桶']
class_en = ['Seashell', 'Lighter','Old Mirror', 'Broom','Ceramic Bowl', 'Toothbrush','Disposable Chopsticks','Dirty Cloth',
            'Newspaper', 'Glassware', 'Basketball', 'Plastic Bottle', 'Cardboard','Glass Bottle', 'Metalware', 'Hats', 'Cans', 'Paper',
            'Vegetable Leaf','Orange Peel', 'Eggshell','Banana Peel',
            'Battery', 'Tablet capsules','Fluorescent lamp', 'Paint bucket']

index_en = {'Seashell': 0, 'Lighter': 1, 'Old Mirror': 2, 'Broom': 3, 'Ceramic Bowl': 4, 'Toothbrush': 5, 'Disposable Chopsticks': 6, 'Dirty Cloth': 7,
            'Newspaper': 8, 'Glassware': 9, 'Basketball': 10, 'Plastic Bottle': 11, 'Cardboard': 12, 'Glass Bottle': 13, 'Metalware': 14, 'Hats': 15, 'Cans': 16, 'Paper': 17,
            'Vegetable Leaf': 18, 'Orange Peel': 19, 'Eggshell': 20, 'Banana Peel': 21,
            'Battery': 22, 'Tablet capsules': 23, 'Fluorescent lamp': 24, 'Paint bucket': 25}

# 训练超参
config = EasyDict({
    "num_classes": 26,
    "image_height": 224,
    "image_width": 224,
    #"data_split": [0.9, 0.1],
    "backbone_out_channels":1280,
    "batch_size": 16,
    "eval_batch_size": 8,
    "epochs": 10,
    "lr_max": 0.05,
    "momentum": 0.9,
    "weight_decay": 1e-4,
    "save_ckpt_epochs": 1,
    "dataset_path": "./data_en",
    "class_index": index_en,
    "pretrained_ckpt": "./mobilenetV2-200_1067.ckpt" # mobilenetV2-200_1067.ckpt 
})
数据预处理操作

利用ImageFolderDataset方法读取垃圾分类数据集,并整体对数据集进行处理。

读取数据集时指定训练集和测试集,首先对整个数据集进行归一化,修改图像频道等预处理操作。然后对训练集的数据依次进行RandomCropDecodeResize、RandomHorizontalFlip、RandomColorAdjust、shuffle操作,以增加训练数据的丰富度;对测试集进行Decode、Resize、CenterCrop等预处理操作;最后返回处理后的数据集。

def create_dataset(dataset_path, config, training=True, buffer_size=1000):
    """
    create a train or eval dataset

    Args:
        dataset_path(string): the path of dataset.
        config(struct): the config of train and eval in diffirent platform.

    Returns:
        train_dataset, val_dataset
    """
    data_path = os.path.join(dataset_path, 'train' if training else 'test')
    ds = de.ImageFolderDataset(data_path, num_parallel_workers=4, class_indexing=config.class_index)
    resize_height = config.image_height
    resize_width = config.image_width
    
    normalize_op = C.Normalize(mean=[0.485*255, 0.456*255, 0.406*255], std=[0.229*255, 0.224*255, 0.225*255])
    change_swap_op = C.HWC2CHW()
    type_cast_op = C2.TypeCast(mstype.int32)

    if training:
        crop_decode_resize = C.RandomCropDecodeResize(resize_height, scale=(0.08, 1.0), ratio=(0.75, 1.333))
        horizontal_flip_op = C.RandomHorizontalFlip(prob=0.5)
        color_adjust = C.RandomColorAdjust(brightness=0.4, contrast=0.4, saturation=0.4)
    
        train_trans = [crop_decode_resize, horizontal_flip_op, color_adjust, normalize_op, change_swap_op]
        train_ds = ds.map(input_columns="image", operations=train_trans, num_parallel_workers=4)
        train_ds = train_ds.map(input_columns="label", operations=type_cast_op, num_parallel_workers=4)
        
        train_ds = train_ds.shuffle(buffer_size=buffer_size)
        ds = train_ds.batch(config.batch_size, drop_remainder=True)
    else:
        decode_op = C.Decode()
        resize_op = C.Resize((int(resize_width/0.875), int(resize_width/0.875)))
        center_crop = C.CenterCrop(resize_width)
        
        eval_trans = [decode_op, resize_op, center_crop, normalize_op, change_swap_op]
        eval_ds = ds.map(input_columns="image", operations=eval_trans, num_parallel_workers=4)
        eval_ds = eval_ds.map(input_columns="label", operations=type_cast_op, num_parallel_workers=4)
        ds = eval_ds.batch(config.eval_batch_size, drop_remainder=True)

    return ds

ds = create_dataset(dataset_path=config.dataset_path, config=config, training=False)
print(ds.get_dataset_size())
data = ds.create_dict_iterator(output_numpy=True)._get_next()
images = data['image']
labels = data['label']

for i in range(1, 5):
    plt.subplot(2, 2, i)
    plt.imshow(np.transpose(images[i], (1,2,0)))
    plt.title('label: %s' % class_en[labels[i]])
    plt.xticks([])
plt.show()

在这里插入图片描述

3、基于MindSpore的MobileNetV2模型搭建

使用MindSpore定义MobileNetV2网络的各模块时需要继承mindspore.nn.Cell。Cell是所有神经网络(Conv2d等)的基类。

神经网络的各层需要预先在__init__方法中定义,然后通过定义construct方法来完成神经网络的前向构造。原始模型激活函数为ReLU6,池化模块采用是全局平均池化层。

ReLU6激活函数

ReLU6(x)=min(max(0,x),6)

ReLU6可以抑制ReLU的最大值,当x>6时,其导数也是0;

使用ReLU6代替ReLU的目的当然是牺牲一定的精度换取轻量化:在移动端float16的低精度的时候,也能有很好的数值分辨率,如果对ReLu的输出值不加限制,那么输出范围就是0到正无穷,而低精度的float16无法精确描述其数值,带来精度损失。

深度可分离卷积

听这名字,就能猜测到相比普通卷积,Depthwise separable convolution 或许既能像残差连接一样增加网络深度防止过早的梯度消失或爆炸,又能减少卷积运算量一样。

深度可分离卷积将一次普通卷积拆分为两个过程,分别为逐通道卷积(Depthwise Convolution)和逐点卷积(Pointwise Convolution)。

  • 逐通道卷积(Depthwise Convolution)
    Depthwise Convolution的一个卷积核负责一个通道,一个通道只被一个卷积核卷积,这个过程产生的feature map通道数和输入的通道数完全一样。

  • 逐点卷积(Pointwise Convolution)
    Pointwise Convolution的运算与常规卷积运算非常相似,它的卷积核的尺寸为 1×1×M,M为上一层的通道数。所以这里的卷积运算会将上一步的map在深度方向上进行加权组合,生成新的Feature map。有几个卷积核就有几个输出Feature map。(卷积核的shape即为:1 x 1 x 输入通道数 x 输出通道数)
    在这里插入图片描述
    用深度可分卷积做两次运算有什么作用?鉴于MobileNetV2是为了加速计算机视觉模型的讨论,那么,如何通过两次操作而不是一次操作来加快速度?

假设我们想对RGB图像应用64个卷积滤波器,使输出中有64个通道。正常卷积层中的参数数量(包括偏置项)为3x3x3x64+64=1792。然而,使用深度可分离卷积时,只有(3x3x1x3+3)+(1x1x64+64)=30+256=286个参数,这是一个显著的减少,深度可分离的卷积的参数小于正常卷积的6倍,同时计算量也会相应降低。

因此,引入深度可分离卷积层将有助于减少计算和参数的数量,从而减少训练/推理时间,并有助于分别正则化我们的模型。模型更轻量化了,搭配算力更强的终端嵌入式设备,性能就会更好。

MobileNetV2网络结构

在MobileNetV2中,有两种类型的块。一种是步幅为1的残差块。另一种是步幅为2的块用于下采样。

这两种类型的块都有3个层。

这次,第一层是1×1卷积,带有ReLU6激活函数。

第二层是深度卷积。

第三层是另一个1×1卷积,但没有任何非线性。据称,如果再次使用ReLU,深度网络只能对输出域的非零体积部分具有线性分类器的能力。
在这里插入图片描述

__all__ = ['MobileNetV2', 'MobileNetV2Backbone', 'MobileNetV2Head', 'mobilenet_v2']

# 在大多数硬件中,size 可以被 d = 8, 16, ... 整除的矩阵乘法比较快,因为这些 size 符合处理器单元的对齐位宽
def _make_divisible(v, divisor, min_value=None):  # _make_divisible可确保所有网络层的通道数都可以被8整除,与硬件设备的优化标准对齐
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    if new_v < 0.9 * v:
        new_v += divisor
    return new_v

class GlobalAvgPooling(nn.Cell):
    """
    Global avg pooling definition.

    Args:

    Returns:
        Tensor, output tensor.

    Examples:
        >>> GlobalAvgPooling()
    """

    def __init__(self):
        super(GlobalAvgPooling, self).__init__()

    def construct(self, x):
        x = P.mean(x, (2, 3))
        return x

class ConvBNReLU(nn.Cell):
    """
    Convolution/Depthwise fused with Batchnorm and ReLU block definition.

    Args:
        in_planes (int): Input channel.
        out_planes (int): Output channel.
        kernel_size (int): Input kernel size.
        stride (int): Stride size for the first convolutional layer. Default: 1.
        groups (int): channel group. Convolution is 1 while Depthiwse is input channel. Default: 1.

    Returns:
        Tensor, output tensor.

    Examples:
        >>> ConvBNReLU(16, 256, kernel_size=1, stride=1, groups=1)
    """

    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
        super(ConvBNReLU, self).__init__()
        padding = (kernel_size - 1) // 2
        in_channels = in_planes
        out_channels = out_planes
        if groups == 1:
            conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, pad_mode='pad', padding=padding)
        else:
            out_channels = in_planes
            conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, pad_mode='pad',
                             padding=padding, group=in_channels)

        layers = [conv, nn.BatchNorm2d(out_planes), nn.ReLU6()]
        self.features = nn.SequentialCell(layers)

    def construct(self, x):
        output = self.features(x)
        return output

class InvertedResidual(nn.Cell):
    """
    Mobilenetv2 residual block definition.

    Args:
        inp (int): Input channel.
        oup (int): Output channel.
        stride (int): Stride size for the first convolutional layer. Default: 1.
        expand_ratio (int): expand ration of input channel

    Returns:
        Tensor, output tensor.

    Examples:
        >>> ResidualBlock(3, 256, 1, 1)
    """

    def __init__(self, inp, oup, stride, expand_ratio):
        super(InvertedResidual, self).__init__()
        assert stride in [1, 2]

        hidden_dim = int(round(inp * expand_ratio))
        self.use_res_connect = stride == 1 and inp == oup  # stride==1且输入通道数==输出通道数才会进行残差相加

        layers = []
        if expand_ratio != 1:
            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
        layers.extend([
            ConvBNReLU(hidden_dim, hidden_dim,
                       stride=stride, groups=hidden_dim),
            nn.Conv2d(hidden_dim, oup, kernel_size=1,
                      stride=1, has_bias=False),
            nn.BatchNorm2d(oup),
        ])
        self.conv = nn.SequentialCell(layers)
        self.cast = P.Cast()

    def construct(self, x):
        identity = x
        x = self.conv(x)
        if self.use_res_connect:
            return P.add(identity, x)
        return x

class MobileNetV2Backbone(nn.Cell):
    """
    MobileNetV2 architecture.

    Args:
        class_num (int): number of classes.
        width_mult (int): Channels multiplier for round to 8/16 and others. Default is 1.
        has_dropout (bool): Is dropout used. Default is false
        inverted_residual_setting (list): Inverted residual settings. Default is None
        round_nearest (list): Channel round to . Default is 8
    Returns:
        Tensor, output tensor.

    Examples:
        >>> MobileNetV2(num_classes=1000)
    """

    def __init__(self, width_mult=1., inverted_residual_setting=None, round_nearest=8,
                 input_channel=32, last_channel=1280):
        super(MobileNetV2Backbone, self).__init__()
        block = InvertedResidual
        # setting of inverted residual blocks
        self.cfgs = inverted_residual_setting
        if inverted_residual_setting is None:
            self.cfgs = [
                # t, c, n, s
                [1, 16, 1, 1],
                [6, 24, 2, 2],
                [6, 32, 3, 2],
                [6, 64, 4, 2],
                [6, 96, 3, 1],
                [6, 160, 3, 2],
                [6, 320, 1, 1],
            ]

        # building first layer
        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
        self.out_channels = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
        features = [ConvBNReLU(3, input_channel, stride=2)]
        # building inverted residual blocks
        for t, c, n, s in self.cfgs:
            output_channel = _make_divisible(c * width_mult, round_nearest)
            for i in range(n):
                stride = s if i == 0 else 1
                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
                input_channel = output_channel
        features.append(ConvBNReLU(input_channel, self.out_channels, kernel_size=1))
        self.features = nn.SequentialCell(features)
        self._initialize_weights()

    def construct(self, x):
        x = self.features(x)
        return x

    def _initialize_weights(self):
        """
        Initialize weights.

        Args:

        Returns:
            None.

        Examples:
            >>> _initialize_weights()
        """
        self.init_parameters_data()
        for _, m in self.cells_and_names():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.set_data(Tensor(np.random.normal(0, np.sqrt(2. / n),
                                                          m.weight.data.shape).astype("float32")))
                if m.bias is not None:
                    m.bias.set_data(
                        Tensor(np.zeros(m.bias.data.shape, dtype="float32")))
            elif isinstance(m, nn.BatchNorm2d):
                m.gamma.set_data(
                    Tensor(np.ones(m.gamma.data.shape, dtype="float32")))
                m.beta.set_data(
                    Tensor(np.zeros(m.beta.data.shape, dtype="float32")))

    @property
    def get_features(self):
        return self.features

class MobileNetV2Head(nn.Cell):
    """
    MobileNetV2 architecture.

    Args:
        class_num (int): Number of classes. Default is 1000.
        has_dropout (bool): Is dropout used. Default is false
    Returns:
        Tensor, output tensor.

    Examples:
        >>> MobileNetV2(num_classes=1000)
    """

    def __init__(self, input_channel=1280, num_classes=1000, has_dropout=False, activation="None"):
        super(MobileNetV2Head, self).__init__()
        # mobilenet head
        head = ([GlobalAvgPooling(), nn.Dense(input_channel, num_classes, has_bias=True)] if not has_dropout else
                [GlobalAvgPooling(), nn.Dropout(0.2), nn.Dense(input_channel, num_classes, has_bias=True)])
        self.head = nn.SequentialCell(head)
        self.need_activation = True
        if activation == "Sigmoid":
            self.activation = nn.Sigmoid()
        elif activation == "Softmax":
            self.activation = nn.Softmax()
        else:
            self.need_activation = False
        self._initialize_weights()

    def construct(self, x):
        x = self.head(x)
        if self.need_activation:
            x = self.activation(x)
        return x

    def _initialize_weights(self):
        """
        Initialize weights.

        Args:

        Returns:
            None.

        Examples:
            >>> _initialize_weights()
        """
        self.init_parameters_data()
        for _, m in self.cells_and_names():
            if isinstance(m, nn.Dense):
                m.weight.set_data(Tensor(np.random.normal(
                    0, 0.01, m.weight.data.shape).astype("float32")))
                if m.bias is not None:
                    m.bias.set_data(
                        Tensor(np.zeros(m.bias.data.shape, dtype="float32")))
    @property
    def get_head(self):
        return self.head

class MobileNetV2(nn.Cell):
    """
    MobileNetV2 architecture.

    Args:
        class_num (int): number of classes.
        width_mult (int): Channels multiplier for round to 8/16 and others. Default is 1.
        has_dropout (bool): Is dropout used. Default is false
        inverted_residual_setting (list): Inverted residual settings. Default is None
        round_nearest (list): Channel round to . Default is 8
    Returns:
        Tensor, output tensor.

    Examples:
        >>> MobileNetV2(backbone, head)
    """

    def __init__(self, num_classes=1000, width_mult=1., has_dropout=False, inverted_residual_setting=None, \
        round_nearest=8, input_channel=32, last_channel=1280):
        super(MobileNetV2, self).__init__()
        self.backbone = MobileNetV2Backbone(width_mult=width_mult, \
            inverted_residual_setting=inverted_residual_setting, \
            round_nearest=round_nearest, input_channel=input_channel, last_channel=last_channel).get_features
        self.head = MobileNetV2Head(input_channel=self.backbone.out_channel, num_classes=num_classes, \
            has_dropout=has_dropout).get_head

    def construct(self, x):
        x = self.backbone(x)
        x = self.head(x)
        return x

class MobileNetV2Combine(nn.Cell):
    """
    MobileNetV2Combine architecture.

    Args:
        backbone (Cell): the features extract layers.
        head (Cell):  the fully connected layers.
    Returns:
        Tensor, output tensor.

    Examples:
        >>> MobileNetV2(num_classes=1000)
    """

    def __init__(self, backbone, head):
        super(MobileNetV2Combine, self).__init__(auto_prefix=False)
        self.backbone = backbone
        self.head = head

    def construct(self, x):
        x = self.backbone(x)
        x = self.head(x)
        return x

def mobilenet_v2(backbone, head):
    return MobileNetV2Combine(backbone, head)
4、模型训练与测试

一般情况下,模型训练时采用静态学习率,如0.01。随着训练步数的增加,模型逐渐趋于收敛,对权重参数的更新幅度应该逐渐降低,以减小模型训练后期的抖动。所以,模型训练时可以采用动态下降的学习率,常见的学习率下降策略有:

  • polynomial decay/square decay;
  • cosine decay;
  • exponential decay;
  • stage decay.

这里使用cosine decay下降策略:

def cosine_decay(total_steps, lr_init=0.0, lr_end=0.0, lr_max=0.1, warmup_steps=0):
    """
    Applies cosine decay to generate learning rate array.

    Args:
       total_steps(int): all steps in training.
       lr_init(float): init learning rate.
       lr_end(float): end learning rate
       lr_max(float): max learning rate.
       warmup_steps(int): all steps in warmup epochs.

    Returns:
       list, learning rate array.
    """
    lr_init, lr_end, lr_max = float(lr_init), float(lr_end), float(lr_max)
    decay_steps = total_steps - warmup_steps
    lr_all_steps = []
    inc_per_step = (lr_max - lr_init) / warmup_steps if warmup_steps else 0
    for i in range(total_steps):
        if i < warmup_steps:
            lr = lr_init + inc_per_step * (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_end
        lr_all_steps.append(lr)

    return lr_all_steps
def switch_precision(net, data_type):
    if ms.get_context('device_target') == "Ascend":
        net.to_float(data_type)
        for _, cell in net.cells_and_names():
            if isinstance(cell, nn.Dense):
                cell.to_float(ms.float32)

在进行正式的训练之前,定义训练函数,读取数据并对模型进行实例化,定义优化器和损失函数。

首先简单介绍损失函数及优化器的概念:

  • 损失函数:也叫成本函数,即深度学习训练的目标函数,用于衡量预测值与实际值差异的程度。深度学习通过不停地迭代来缩小损失函数的值。定义一个好的损失函数,可以有效提高模型的性能。

  • 优化器:用于最小化损失函数,从而在训练过程中改进模型。

定义了损失函数后,可以得到损失函数关于权重的梯度。梯度用于指示优化器优化权重的方向,以提高模型性能。

在训练MobileNetV2之前对MobileNetV2Backbone层的参数进行了固定,使其在训练过程中对该模块的权重参数不进行更新;只对MobileNetV2Head模块的参数进行更新。

MindSpore支持的损失函数有SoftmaxCrossEntropyWithLogits、L1Loss、MSELoss等。图像分类这里使用SoftmaxCrossEntropyWithLogits损失函数。

训练测试过程中会打印loss值,loss值会波动,但总体来说loss值会逐步减小,精度逐步提高。每个人运行的loss值有一定随机性,不一定完全相同。

每打印一个epoch后模型都会在测试集上的计算测试精度,从打印的精度值分析MobileNetV2模型的预测能力在不断提升。

在模型训练过程中,可以添加检查点(Checkpoint)用于保存模型的参数,以便进行推理及中断后再训练使用。使用场景如下:

  • 训练后推理场景
  1. 模型训练完毕后保存模型的参数,用于推理或预测操作。
  2. 训练过程中,通过实时验证精度,把精度最高的模型参数保存下来,用于预测操作。
  • 再训练场景
  1. 进行长时间训练任务时,保存训练过程中的Checkpoint文件,防止任务异常退出后从初始状态开始训练。
  2. Fine-tuning(微调)场景,即训练一个模型并保存参数,基于该模型,面向第二个类似任务进行模型训练。

这里加载ImageNet数据上预训练的MobileNetv2进行Fine-tuning,只训练最后修改的FC层,并在训练过程中保存Checkpoint。

from mindspore.amp import FixedLossScaleManager
import time
LOSS_SCALE = 1024

train_dataset = create_dataset(dataset_path=config.dataset_path, config=config)
eval_dataset = create_dataset(dataset_path=config.dataset_path, config=config)
step_size = train_dataset.get_dataset_size()
    
backbone = MobileNetV2Backbone() #last_channel=config.backbone_out_channels
# Freeze parameters of backbone. You can comment these two lines.
for param in backbone.get_parameters():
    param.requires_grad = False
# load parameters from pretrained model
load_checkpoint(config.pretrained_ckpt, backbone)

head = MobileNetV2Head(input_channel=backbone.out_channels, num_classes=config.num_classes)
network = mobilenet_v2(backbone, head)

# define loss, optimizer, and model
loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
loss_scale = FixedLossScaleManager(LOSS_SCALE, drop_overflow_update=False)
lrs = cosine_decay(config.epochs * step_size, lr_max=config.lr_max)
opt = nn.Momentum(network.trainable_params(), lrs, config.momentum, config.weight_decay, loss_scale=LOSS_SCALE)

# 定义用于训练的train_loop函数。
def train_loop(model, dataset, loss_fn, optimizer):
    # 定义正向计算函数
    def forward_fn(data, label):
        logits = model(data)
        loss = loss_fn(logits, label)
        return loss

    # 定义微分函数,使用mindspore.value_and_grad获得微分函数grad_fn,输出loss和梯度。
    # 由于是对模型参数求导,grad_position 配置为None,传入可训练参数。
    grad_fn = ms.value_and_grad(forward_fn, None, optimizer.parameters)

    # 定义 one-step training函数
    def train_step(data, label):
        loss, grads = grad_fn(data, label)
        optimizer(grads)
        return loss

    size = dataset.get_dataset_size()
    model.set_train()
    for batch, (data, label) in enumerate(dataset.create_tuple_iterator()):
        loss = train_step(data, label)

        if batch % 10 == 0:
            loss, current = loss.asnumpy(), batch
            print(f"loss: {loss:>7f}  [{current:>3d}/{size:>3d}]")

# 定义用于测试的test_loop函数。
def test_loop(model, dataset, loss_fn):
    num_batches = dataset.get_dataset_size()
    model.set_train(False)
    total, test_loss, correct = 0, 0, 0
    for data, label in dataset.create_tuple_iterator():
        pred = model(data)
        total += len(data)
        test_loss += loss_fn(pred, label).asnumpy()
        correct += (pred.argmax(1) == label).asnumpy().sum()
    test_loss /= num_batches
    correct /= total
    print(f"Test: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")

print("============== Starting Training ==============")
# 由于时间问题,训练过程只进行了2个epoch ,可以根据需求调整。
epoch_begin_time = time.time()
epochs = 2
for t in range(epochs):
    begin_time = time.time()
    print(f"Epoch {t+1}\n-------------------------------")
    train_loop(network, train_dataset, loss, opt)
    ms.save_checkpoint(network, "save_mobilenetV2_model.ckpt")
    end_time = time.time()
    times = end_time - begin_time
    print(f"per epoch time: {times}s")
    test_loop(network, eval_dataset, loss)
epoch_end_time = time.time()
times = epoch_end_time - epoch_begin_time
print(f"total time:  {times}s")
print("============== Training Success ==============")

在这里插入图片描述

5、模型推理

加载模型Checkpoint进行推理,使用load_checkpoint接口加载数据时,需要把数据传入给原始网络,而不能传递给带有优化器和损失函数的训练网络。

CKPT="save_mobilenetV2_model.ckpt"
def image_process(image):
    """Precess one image per time.
    
    Args:
        image: shape (H, W, C)
    """
    mean=[0.485*255, 0.456*255, 0.406*255]
    std=[0.229*255, 0.224*255, 0.225*255]
    image = (np.array(image) - mean) / std
    image = image.transpose((2,0,1))
    img_tensor = Tensor(np.array([image], np.float32))
    return img_tensor

def infer_one(network, image_path):
    image = Image.open(image_path).resize((config.image_height, config.image_width))
    logits = network(image_process(image))
    pred = np.argmax(logits.asnumpy(), axis=1)[0]
    print(image_path, class_en[pred])

def infer():
    backbone = MobileNetV2Backbone(last_channel=config.backbone_out_channels)
    head = MobileNetV2Head(input_channel=backbone.out_channels, num_classes=config.num_classes)
    network = mobilenet_v2(backbone, head)
    load_checkpoint(CKPT, network)
    for i in range(91, 100):
        infer_one(network, f'data_en/test/Cardboard/000{i}.jpg')  # 10张Cardboard测试图片
infer()

#print_log:
data_en/test/Cardboard/00091.jpg Battery
data_en/test/Cardboard/00092.jpg Ceramic Bowl
data_en/test/Cardboard/00093.jpg Glassware
data_en/test/Cardboard/00094.jpg Cardboard
data_en/test/Cardboard/00095.jpg Paper
data_en/test/Cardboard/00096.jpg Toothbrush
data_en/test/Cardboard/00097.jpg Broom
data_en/test/Cardboard/00098.jpg Cardboard
data_en/test/Cardboard/00099.jpg Glass Bottle

Reference

[1] 昇思大模型平台
[2] 昇思官方文档-基于MobileNetv2的垃圾分类
[3] MobileNetV2 — 轻量级模型(图像分类)
[4] ReLU6
[5] 深度可分离卷积(Depthwise seperable convolution)
[6] 深度可分离卷积实战


原文地址:https://blog.csdn.net/wayne6515/article/details/140710539

免责声明:本站文章内容转载自网络资源,如本站内容侵犯了原著者的合法权益,可联系本站删除。更多内容请关注自学内容网(zxcms.com)!