mask2former 语义分割 自定义数据集 语义分割代码实现

文章目录

• 前言
• 语义分割发展史及意义
• 一、数据集的准备
• 二、基于MegEngine的语义分割框架构建

• 1.引入库
• 2.CPU/GPU配置
• 3.数据标准化
• 4.解析数据集到列表中
• 5.设置数据迭代器
• 6.获取loader
• 7.模型构建
• 8.模型训练

• 1.优化器及参数初始化
• 2.模型训练
• 3.模型保存

• 9.模型预测

• 三、基于MegEngine的模型构建
• 四、模型主入口
• 五、效果展示

• 1.自定义数据集效果展示
• 2.VOC数据集效果展示

• 六、各模型免费下载
• 总结

---

前言

  本文主要讲解基于megengine深度学习框架实现目标检测，鉴于之前写chainer的麻烦，本结构代码也类似chainer的目标检测框架，各个模型只需要修改网络结构即可，本次直接一篇博文写完目标检测框架及网络结构的搭建，让志同道合者不需要切换文章。

环境配置：
      python 3.8
      megengine 1.9.1
      cuda 10.1

---

语义分割发展史及意义

  图像语义分割，它是将整个图像分成一个个像素组，然后对其进行标记和分类。特别地，语义分割试图在语义上理解图像中每个像素的角色。

  图像语义分割可以说是图像理解的基石性技术，在自动驾驶系统中举足轻重。众所周知，图像是由一个个像素（Pixel）组成的，而语义分割就是将图像中表达语义含义的不同进行分组(Grouping)/分割(Segmentation)。语义图像分割就是将每个像素都标注上其对应的类别。需要注意的是，这里不但单独区分同一类别的不同个体，而是仅仅关系该像素是属于哪个类别。

  分割任务对于许多任务都非常有用，比如自动驾驶汽车(为了使自动驾驶汽车能够适应现存道路，其需要具有对周围环境的感知能力)；医疗图像判断(可以通过机器辅助放射治疗师的分析，从而加速放射检查)

---

一、数据集的准备

语义分割数据标注主要使用labelImg工具，python安装只需要：pip install labelme 即可，然后在命令提示符输入：labelme即可，如图：

在这里只需要修改“OpenDir“，“OpenDir“主要是存放图片需要标注的路径

选择好路径之后即可开始绘制：

我在平时标注的时候快捷键一般只用到：

createpolygons：（ctrl+N）开始绘制

a：上一张

d：下一张

绘制过程如图：

就只需要一次把目标绘制完成即可。

---

二、基于MegEngine的语义分割框架构建

本语义分割架目录结构如下：

BaseModel：此目录保存基于VOC数据集、cityspace数据集已经训练好的模型

core：此目录主要保存标准的py文件，功能如语义分割评估算法等计算

data：此目录主要保存标准py文件，功能如数据加载器，迭代器等

nets：此目录主要保存模型结构

result_Model：此目录是自定义训练数据集的保存模型位置

Ctu_Segmentation.py：语义分割主类实现及主入口

1.引入库

import os,time,megengine,math,json,random,cv2,sys
sys.path.append('.')
import numpy as np
from PIL import Image
import megengine as mge
import megengine.distributed as dist
import megengine.functional as F
from megengine.autodiff import GradManager
from megengine.data import DataLoader, Infinite, RandomSampler
from megengine.data import transform as T
from megengine.optimizer import SGD

2.CPU/GPU配置

if USEGPU!='-1' and  dist.helper.get_device_count_by_fork("gpu") > 0:
    megengine.set_default_device('gpux')
else:
    megengine.set_default_device('cpux')
    USEGPU='-1'
os.environ['CUDA_VISIBLE_DEVICES']= USEGPU

3.数据标准化

self.img_mean = [103.530, 116.280, 123.675]  # BGR
self.img_std = [57.375, 57.120, 58.395]

4.解析数据集到列表中

def CreateDataList_Segmentation(DataDir,train_split=0.9):
    DataList = [os.path.join(DataDir, fileEach) for fileEach in [each for each in os.listdir(DataDir) if re.match(r'.*\.json', each)]]
    All_label_name = ['_background_']
    for each_json in DataList:
        try:
            data = json.load(open(each_json))
        except:
            try:
                data = json.load(open(each_json,encoding='utf-8'))
            except:
                data = json.load(open(each_json,encoding='utf-8',errors='ignore'))
        for shape in sorted(data['shapes'], key=lambda x: x['label']):
            label_name = shape['label']
            if label_name not in All_label_name:
                All_label_name.append(label_name)
    random.shuffle(DataList)
    if train_split<=0 or train_split>=1:
        train_list = DataList
        val_list = DataList
    else:
        train_list = DataList[:int(len(DataList)*train_split)]
        val_list = DataList[int(len(DataList)*train_split):]
    return train_list,val_list, All_label_name

def JsonToLabel(JsonFile,classes_names):
    try:
        data = json.load(open(JsonFile))
    except:
        try:
            data = json.load(open(JsonFile,encoding='utf-8'))
        except:
            data = json.load(open(JsonFile,encoding='utf-8',errors='ignore'))
    imageData = data.get('imageData')

    label_name_to_value = {'_background_': 0}
    for shape in sorted(data['shapes'], key=lambda x: x['label']):
        label_name = shape['label']
        if label_name not in label_name_to_value:
            label_value = len(label_name_to_value)
            label_name_to_value[label_name] = label_value

    img = utils.img_b64_to_arr(imageData)
    lbl, _ = utils.shapes_to_label(img.shape, data['shapes'], label_name_to_value)
    lbl_pil = PIL.Image.fromarray(lbl.astype(np.uint8), mode='P')
    colormap = imgviz.label_colormap()
    lbl_pil.putpalette(colormap.flatten())
    label_names = [None] * (max(label_name_to_value.values()) + 1)
    for name, value in label_name_to_value.items():
        label_names[value] = name

    new = Image.new("RGB", [np.shape(lbl_pil)[1], np.shape(lbl_pil)[0]])
    for name in label_names:
        index_json = label_names.index(name)
        index_all = classes_names.index(name)
        new = new + np.expand_dims(index_all * (np.array(lbl_pil) == index_json), -1)
    new = Image.fromarray(np.uint8(new))
    lab = cv2.cvtColor(np.asarray(new), cv2.COLOR_RGB2GRAY)
    img = cv2.cvtColor(np.asarray(img), cv2.COLOR_RGB2BGR)

    return img, lab

5.设置数据迭代器

class PascalVOC(Dataset):
    def __init__(self, data_list,classes_names):
        super().__init__()
        self.data_list = data_list
        self.classes_names = classes_names
        self.img_infos = dict()

    def __getitem__(self, index):
        target = []
        image,mask = JsonToLabel(self.data_list[index],self.classes_names)
        target.append(image)
        mask = mask[:, :, np.newaxis]
        target.append(mask)
        return tuple(target)

    def __len__(self):
        return len(self.data_list)

6.获取loader

self.train_dataloader = DataLoader(
    train_dataset,
    sampler=train_sampler,
    transform=T.Compose(
        transforms=[
            T.RandomHorizontalFlip(0.5),
            T.RandomResize(scale_range=(0.5, 2)),
            T.RandomCrop(
                output_size=(self.image_size, self.image_size),
                padding_value=[0, 0, 0],
                padding_maskvalue=255,
            ),
            T.Normalize(mean=self.img_mean, std=self.img_std),
            T.ToMode(),
        ],
        order=["image", "mask"],
    ),
    num_workers=num_workers,
)

7.模型构建

self.model = DeepLabV3Plus(num_classes=len(self.classes_names),backbone=backbone)

8.模型训练

1.优化器及参数初始化

backbone_params = []
head_params = []
for name, param in self.model.named_parameters():
    if "backbone" in name:
        backbone_params.append(param)
    else:
        head_params.append(param)

opt = SGD(
    [
        {
            "params": backbone_params,
            "lr": learning_rate  * 0.1,
        },
        {"params": head_params},
    ],
    lr=learning_rate,
    momentum=0.9,
    weight_decay=0.0001,
)
gm = GradManager()
gm.attach(self.model.parameters())

2.模型训练

with gm:
    pred = self.model(mge.tensor(inputs))
    mask = mge.tensor(labels) != 255
    pred = pred.transpose(0, 2, 3, 1)
    loss =  F.loss.cross_entropy(pred[mask], mge.tensor(labels)[mask], 1)
    gm.backward(loss)
opt.step().clear_grad()

3.模型保存

模型保存结构必须保持结构一致，如图：

主要包含主干网络结构，类别，图像大小，模型保存路径

megengine.save(
     {"epoch": epoch, "state_dict": self.model.state_dict()}, ClassDict['model_path'],
)

9.模型预测

def predict(self,img_cv):
    ori_h, ori_w = img_cv.shape[:2]
    processed_img = self.transform.apply(img_cv)[np.newaxis, :]
    processed_img = megengine.tensor(processed_img, dtype="float32")
    pred = self.model(processed_img)
    pred = pred.numpy().squeeze().argmax(0)
    pred = cv2.resize(pred.astype("uint8"), (ori_w, ori_h), interpolation=cv2.INTER_NEAREST)

    seg_img = np.zeros((ori_h, ori_w, 3))
    for c in range(len(self.classes_names)):
        seg_img[:,:,0] += ((pred[:,: ] == c )*( self.colors[c][0] )).astype('uint8')
        seg_img[:,:,1] += ((pred[:,: ] == c )*( self.colors[c][1] )).astype('uint8')
        seg_img[:,:,2] += ((pred[:,: ] == c )*( self.colors[c][2] )).astype('uint8')
    image_result = seg_img.astype("uint8")
    img_add = cv2.addWeighted(img_cv, 1.0, image_result, 0.7, 0)
    return img_add

---

三、基于MegEngine的模型构建

deeplab3+代码实现

import megengine.functional as F
import megengine.module as M
from nets.resnet import resnet18,resnet34,resnet50,resnet101,resnet152,resnext50_32x4d,resnext101_32x8d

class ASPP(M.Module):
    def __init__(self, in_channels, out_channels, dr=1):
        super().__init__()

        self.conv1 = M.Sequential(
            M.Conv2d(
                in_channels, out_channels, 1, 1, padding=0, dilation=dr, bias=False
            ),
            M.BatchNorm2d(out_channels),
            M.ReLU(),
        )
        self.conv2 = M.Sequential(
            M.Conv2d(
                in_channels,
                out_channels,
                3,
                1,
                padding=6 * dr,
                dilation=6 * dr,
                bias=False,
            ),
            M.BatchNorm2d(out_channels),
            M.ReLU(),
        )
        self.conv3 = M.Sequential(
            M.Conv2d(
                in_channels,
                out_channels,
                3,
                1,
                padding=12 * dr,
                dilation=12 * dr,
                bias=False,
            ),
            M.BatchNorm2d(out_channels),
            M.ReLU(),
        )
        self.conv4 = M.Sequential(
            M.Conv2d(
                in_channels,
                out_channels,
                3,
                1,
                padding=18 * dr,
                dilation=18 * dr,
                bias=False,
            ),
            M.BatchNorm2d(out_channels),
            M.ReLU(),
        )
        self.conv_gp = M.Sequential(
            M.Conv2d(in_channels, out_channels, 1, 1, 0, bias=False),
            M.BatchNorm2d(out_channels),
            M.ReLU(),
        )
        self.conv_out = M.Sequential(
            M.Conv2d(out_channels * 5, out_channels, 1, 1, padding=0, bias=False),
            M.BatchNorm2d(out_channels),
            M.ReLU(),
        )

    def forward(self, x):
        conv1 = self.conv1(x)
        conv31 = self.conv2(x)
        conv32 = self.conv3(x)
        conv33 = self.conv4(x)

        gp = F.mean(x, [2, 3], True)
        gp = self.conv_gp(gp)
        gp = F.nn.interpolate(gp, x.shape[2:])

        out = F.concat([conv1, conv31, conv32, conv33, gp], axis=1)
        out = self.conv_out(out)
        return out

class DeepLabV3Plus(M.Module):
    def __init__(self, num_classes=21,backbone='resnet18'):
        super().__init__()

        self.output_stride = 16
        self.num_classes = num_classes
        self.model_name={
            'resnet18':resnet18,
            'resnet34':resnet34,
            'resnet50':resnet50,
            'resnet101':resnet101,
            'resnet152':resnet152,
            'resnext50_32x4d':resnext50_32x4d,
            'resnext101_32x8d':resnext101_32x8d
        }
        self.backbone_net=backbone

        self.aspp = ASPP(
            in_channels=2048, out_channels=256, dr=16 // self.output_stride
        )
        self.dropout = M.Dropout(0.5)

        self.upstage1 = M.Sequential(
            M.Conv2d(256, 48, 1, 1, padding=0, bias=False),
            M.BatchNorm2d(48),
            M.ReLU(),
        )

        self.upstage2 = M.Sequential(
            M.Conv2d(256 + 48, 256, 3, 1, padding=1, bias=False),
            M.BatchNorm2d(256),
            M.ReLU(),
            M.Dropout(0.5),
            M.Conv2d(256, 256, 3, 1, padding=1, bias=False),
            M.BatchNorm2d(256),
            M.ReLU(),
            M.Dropout(0.1),
        )
        self.conv_out = M.Conv2d(256, self.num_classes, 1, 1, padding=0)

        for m in self.modules():
            if isinstance(m, M.Conv2d):
                M.init.msra_normal_(m.weight, mode="fan_out", nonlinearity="relu")
            elif isinstance(m, M.BatchNorm2d):
                M.init.ones_(m.weight)
                M.init.zeros_(m.bias)

        if self.backbone_net == 'resnet18' or self.backbone_net=='resnet34':
            self.backbone = self.model_name[self.backbone_net](
                replace_stride_with_dilation=[False, False, False]
            )
        else:
            self.backbone = self.model_name[self.backbone_net](
                replace_stride_with_dilation=[False, False, True]
            )
        del self.backbone.fc

    def forward(self, x):
        layers = self.backbone.extract_features(x)

        up0 = self.aspp(layers["res5"])
        up0 = self.dropout(up0)
        up0 = F.nn.interpolate(up0, layers["res2"].shape[2:])

        up1 = self.upstage1(layers["res2"])
        up1 = F.concat([up0, up1], 1)

        up2 = self.upstage2(up1)

        out = self.conv_out(up2)
        out = F.nn.interpolate(out, x.shape[2:])
        return out

---

四、模型主入口

本人主要习惯简洁方式，因此主入口的实现过程也是很简单的

if __name__ == "__main__":
    # ctu = Ctu_Segmentation(USEGPU="0",image_size=512)
    # ctu.InitModel(r'E:\DL_Project\DataSet\DataSet_Segmentation\DataSet_AQ_HuaHeng1\DataImage',train_split=1,batch_size=2,Pre_Model=None,backbone='resnet50',num_workers=0)
    # ctu.train(TrainNum=400,learning_rate=0.0001, ModelPath='result_Model')

    ctu = Ctu_Segmentation(USEGPU="-1")
    ctu.LoadModel("./BaseModel/ctu_params_voc.json")
    for root, dirs, files in os.walk(r'./BaseModel'):
        for f in files:
            img_cv = ctu.read_image(os.path.join(root, f))
            if img_cv is None:
                continue
            res_img = ctu.predict(img_cv)
            cv2.imshow("result", res_img)
            cv2.waitKey()

---

五、效果展示

1.自定义数据集效果展示

待定

2.VOC数据集效果展示

---