DeepLabV3+语义分割实战

语义分割是计算机视觉的一项重要任务，本文使用Jittor框架实现了DeepLabV3+语义分割模型。

DeepLabV3+论文：https://arxiv.org/pdf/1802.02611.pdf

完整代码：https://github.com/Jittor/deeplab-jittor

1. 数据集

1.1 数据准备

VOC2012数据集是目标检测、语义分割等任务常用的数据集之一， 本文使用VOC数据集的2012 trainaug (train + sbd set)作为训练集，2012 val set作为测试集。

VOC数据集中的物体共包括20个前景类别：'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor' 和背景类别

 

 最终数据集的文件组织如下。

# 文件组织

根目录

|----voc_aug

|    |----datalist

|    |    |----train.txt

|    |    |----val.txt

|    |----images

|    |----annotations

1.2 数据加载

使用jittor.dataset.dataset的基类Dataset可以构造自己的数据集，需要实现__init__、__getitem__、函数。

1. __init__: 定义数据路径，这里的data_root需设置为之前设定的 voc_aug, split 为 train val test 之一，表示选择训练集、验证集还是测试集。同时需要调用self.set_attr来指定数据集加载所需的参数batch_size，total_len、shuffle。
2. __getitem__: 返回单个item的数据。

import numpy as np

import os

from PIL import Image

import matplotlib.pyplot as plt

from jittor.dataset.dataset import Dataset, dataset_root

import jittor as jt

import os

import os.path as osp

from PIL import Image, ImageOps, ImageFilter

import numpy as np

import scipy.io as sio

import random

def fetch(image_path, label_path):

    with open(image_path, 'rb') as fp:

        image = Image.open(fp).convert('RGB')

    with open(label_path, 'rb') as fp:

        label = Image.open(fp).convert('P')

    return image, label

def scale(image, label):

    SCALES = (0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0)

    ratio = np.random.choice(SCALES)

    w,h = image.size

    nw = (int)(w*ratio)

    nh = (int)(h*ratio)

    image = image.resize((nw, nh), Image.BILINEAR)

    label = label.resize((nw, nh), Image.NEAREST)

    return image, label

def pad(image, label):

    w,h = image.size

    crop_size = 513

    pad_h = max(crop_size - h, 0)

    pad_w = max(crop_size - w, 0)

    image = ImageOps.expand(image, border=(0, 0, pad_w, pad_h), fill=0)

    label = ImageOps.expand(label, border=(0, 0, pad_w, pad_h), fill=255)

    return image, label

def crop(image, label):

    w, h = image.size

    crop_size = 513

    x1 = random.randint(0, w - crop_size)

    y1 = random.randint(0, h - crop_size)

    image = image.crop((x1, y1, x1 + crop_size, y1 + crop_size))

    label = label.crop((x1, y1, x1 + crop_size, y1 + crop_size))

    return image, label

def normalize(image, label):

    mean = (0.485, 0.456, 0.40)

    std = (0.229, 0.224, 0.225)

    image = np.array(image).astype(np.float32)

    label = np.array(label).astype(np.float32)

    image /= 255.0

    image -= mean

    image /= std

    return image, label

def flip(image, label):

    if random.random() < 0.5:

        image = image.transpose(Image.FLIP_LEFT_RIGHT)

        label = label.transpose(Image.FLIP_LEFT_RIGHT)

    return image, label

class BaseDataset(Dataset):

    def __init__(self,  data_root='/voc/', split='train', batch_size=1, shuffle=False):

        super().__init__()

        ''' total_len , batch_size, shuffle must be set '''

        self.data_root = data_root

        self.split = split

        self.batch_size = batch_size

        self.shuffle = shuffle

        self.image_root = os.path.join(data_root, 'images')

        self.label_root = os.path.join(data_root, 'annotations')

        self.data_list_path = os.path.join(self.data_root,'/datalist/' + self.split + '.txt')

        self.image_path = []

        self.label_path = []

        with open(self.data_list_path, "r") as f:

            lines = f.read().splitlines()

        for idx, line in enumerate(lines):

            _img_path = os.path.join(self.image_root, line + '.jpg')

            _label_path = os.path.join(self.label_root, line + '.png')

            assert os.path.isfile(_img_path)

            assert os.path.isfile(_label_path)

            self.image_path.append(_img_path)

            self.label_path.append(_label_path)

        self.total_len = len(self.image_path)

        # set_attr must be called to set batch size total len and shuffle like __len__ function in pytorch

        self.set_attr(batch_size = self.batch_size, total_len = self.total_len, shuffle = self.shuffle) # bs , total_len, shuffle

    def __getitem__(self, image_id):

        return NotImplementedError

class TrainDataset(BaseDataset):

    def __init__(self,  data_root='/voc/', split='train', batch_size=1, shuffle=False):

        super(TrainDataset, self).__init__(data_root, split, batch_size, shuffle)

    def __getitem__(self, image_id):

        image_path = self.image_path[image_id]

        label_path = self.label_path[image_id]

        image, label = fetch(image_path, label_path)

        image, label = scale(image, label)

        image, label = pad(image, label)

        image, label = crop(image, label)

        image, label = flip(image, label)

        image, label = normalize(image, label)

        image = np.array(image).astype(np.float).transpose(2, 0, 1)

        image = jt.array(image)

        label = jt.array(np.array(label).astype(np.int))

        return image, label

class ValDataset(BaseDataset):

    def __init__(self,  data_root='/voc/', split='train', batch_size=1, shuffle=False):

        super(ValDataset, self).__init__(data_root, split, batch_size, shuffle)

    def __getitem__(self, image_id):

        image_path = self.image_path[image_id]

        label_path = self.label_path[image_id]

        image, label = fetch(image_path, label_path)

        image, label = normalize(image, label)

        image = np.array(image).astype(np.float).transpose(2, 0, 1)

        image = jt.array(image)

        label = jt.array(np.array(label).astype(np.int))

        return image, label

2. 模型定义

 

 上图为DeepLabV3+论文给出的网络架构图。本文采用ResNe为backbone。输入图像尺寸为513*513。

整个网络可以分成 backbone aspp decoder 三个部分。

2.1 backbonb 这里使用最常见的ResNet，作为backbone并且在ResNet的最后两次使用空洞卷积来扩大感受野，其完整定义如下：

import jittor as jt

from jittor import nn

from jittor import Module

from jittor import init

from jittor.contrib import concat, argmax_pool

import time

class Bottleneck(Module):

    expansion = 4

    def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None):

        super(Bottleneck, self).__init__()

        self.conv1 = nn.Conv(inplanes, planes, kernel_size=1, bias=False)

        self.bn1 = nn.BatchNorm(planes)

        self.conv2 = nn.Conv(planes, planes, kernel_size=3, stride=stride,

                               dilation=dilation, padding=dilation, bias=False)

        self.bn2 = nn.BatchNorm(planes)

        self.conv3 = nn.Conv(planes, planes * 4, kernel_size=1, bias=False)

        self.bn3 = nn.BatchNorm(planes * 4)

        self.relu = nn.ReLU()

        self.downsample = downsample

        self.stride = stride

        self.dilation = dilation

    def execute(self, x):

        residual = x

        out = self.conv1(x)

        out = self.bn1(out)

        out = self.relu(out)

        out = self.conv2(out)

        out = self.bn2(out)

        out = self.relu(out)

        out = self.conv3(out)

        out = self.bn3(out)

        if self.downsample is not None:

            residual = self.downsample(x)

        out += residual

        out = self.relu(out)

        return out

class ResNet(Module):

    def __init__(self, block, layers, output_stride):

        super(ResNet, self).__init__()

        self.inplanes = 64

        blocks = [1, 2, 4]

        if output_stride == 16:

            strides = [1, 2, 2, 1]

            dilations = [1, 1, 1, 2]

        elif output_stride == 8:

            strides = [1, 2, 1, 1]

            dilations = [1, 1, 2, 4]

        else:

            raise NotImplementedError

        # Modules

        self.conv1 = nn.Conv(3, 64, kernel_size=7, stride=2, padding=3, bias=False)

        self.bn1 = nn.BatchNorm(64)

        self.relu = nn.ReLU()

        # self.maxpool = nn.Pool(kernel_size=3, stride=2, padding=1)

        self.layer1 = self._make_layer(block, 64, layers[0], stride=strides[0], dilation=dilations[0])

        self.layer2 = self._make_layer(block, 128, layers[1], stride=strides[1], dilation=dilations[1])

        self.layer3 = self._make_layer(block, 256, layers[2], stride=strides[2], dilation=dilations[2])

        self.layer4 = self._make_MG_unit(block, 512, blocks=blocks, stride=strides[3], dilation=dilations[3])

    def _make_layer(self, block, planes, blocks, stride=1, dilation=1):

        downsample = None

        if stride != 1 or self.inplanes != planes * block.expansion:

            downsample = nn.Sequential(

                nn.Conv(self.inplanes, planes * block.expansion,

                          kernel_size=1, stride=stride, bias=False),

                nn.BatchNorm(planes * block.expansion),

            )

        layers = []

        layers.append(block(self.inplanes, planes, stride, dilation, downsample))

        self.inplanes = planes * block.expansion

        for i in range(1, blocks):

            layers.append(block(self.inplanes, planes, dilation=dilation))

        return nn.Sequential(*layers)

    def _make_MG_unit(self, block, planes, blocks, stride=1, dilation=1):

        downsample = None

        if stride != 1 or self.inplanes != planes * block.expansion:

            downsample = nn.Sequential(

                nn.Conv(self.inplanes, planes * block.expansion,

                          kernel_size=1, stride=stride, bias=False),

                nn.BatchNorm(planes * block.expansion),

            )

        layers = []

        layers.append(block(self.inplanes, planes, stride, dilation=blocks[0]*dilation,

                            downsample=downsample))

        self.inplanes = planes * block.expansion

        for i in range(1, len(blocks)):

            layers.append(block(self.inplanes, planes, stride=1,

                                dilation=blocks[i]*dilation))

        return nn.Sequential(*layers)

    def execute(self, input):

        x = self.conv1(input)

        x = self.bn1(x)

        x = self.relu(x)

        x = argmax_pool(x, 2, 2)

        x = self.layer1(x)

        low_level_feat = x

        x = self.layer2(x)

        x = self.layer3(x)

        x = self.layer4(x)

        return x, low_level_feat

def resnet50(output_stride):

    model = ResNet(Bottleneck, [3,4,6,3], output_stride)

    return model

def resnet101(output_stride):

    model = ResNet(Bottleneck, [3,4,23,3], output_stride)

    return model

2.2 ASPP   

即使用不同尺寸的 dilation conv 对 backbone 得到的 feature map 进行卷积，最后 concat 并整合得到新的特征。

import jittor as jt

from jittor import nn

from jittor import Module

from jittor import init

from jittor.contrib import concat

class Single_ASPPModule(Module):

    def __init__(self, inplanes, planes, kernel_size, padding, dilation):

        super(Single_ASPPModule, self).__init__()

        self.atrous_conv = nn.Conv(inplanes, planes, kernel_size=kernel_size,

                                            stride=1, padding=padding, dilation=dilation, bias=False)

        self.bn = nn.BatchNorm(planes)

        self.relu = nn.ReLU()

    def execute(self, x):

        x = self.atrous_conv(x)

        x = self.bn(x)

        x = self.relu(x)

        return x

class ASPP(Module):

    def __init__(self, output_stride):

        super(ASPP, self).__init__()

        inplanes = 2048

        if output_stride == 16:

            dilations = [1, 6, 12, 18]

        elif output_stride == 8:

            dilations = [1, 12, 24, 36]

        else:

            raise NotImplementedError

        self.aspp1 = Single_ASPPModule(inplanes, 256, 1, padding=0, dilation=dilations[0])

        self.aspp2 = Single_ASPPModule(inplanes, 256, 3, padding=dilations[1], dilation=dilations[1])

        self.aspp3 = Single_ASPPModule(inplanes, 256, 3, padding=dilations[2], dilation=dilations[2])

        self.aspp4 = Single_ASPPModule(inplanes, 256, 3, padding=dilations[3], dilation=dilations[3])

        self.global_avg_pool = nn.Sequential(GlobalPooling(),

                                             nn.Conv(inplanes, 256, 1, stride=1, bias=False),

                                             nn.BatchNorm(256),

                                             nn.ReLU())

        self.conv1 = nn.Conv(1280, 256, 1, bias=False)

        self.bn1 = nn.BatchNorm(256)

        self.relu = nn.ReLU()

        self.dropout = nn.Dropout(0.5)

    def execute(self, x):

        x1 = self.aspp1(x)

        x2 = self.aspp2(x)

        x3 = self.aspp3(x)

        x4 = self.aspp4(x)

        x5 = self.global_avg_pool(x)

        x5 = x5.broadcast((1,1,x4.shape[2],x4.shape[3]))

        x = concat((x1, x2, x3, x4, x5), dim=1)

        x = self.conv1(x)

        x = self.bn1(x)

        x = self.relu(x)

        x = self.dropout(x)

        return x

class GlobalPooling (Module):

    def __init__(self):

        super(GlobalPooling, self).__init__()

    def execute (self, x):

        return jt.mean(x, dims=[2,3], keepdims=1)

2.3 Decoder:

Decoder 将 ASPP 的特征放大后与 ResNet 的中间特征一起 concat， 得到最后分割所用的特征。

import jittor as jt

from jittor import nn

from jittor import Module

from jittor import init

from jittor.contrib import concat

import time

class Decoder(nn.Module):

    def __init__(self, num_classes):

        super(Decoder, self).__init__()

        low_level_inplanes = 256

        self.conv1 = nn.Conv(low_level_inplanes, 48, 1, bias=False)

        self.bn1 = nn.BatchNorm(48)

        self.relu = nn.ReLU()

        self.last_conv = nn.Sequential(nn.Conv(304, 256, kernel_size=3, stride=1, padding=1, bias=False),

                                       nn.BatchNorm(256),

                                       nn.ReLU(),

                                       nn.Dropout(0.5),

                                       nn.Conv(256, 256, kernel_size=3, stride=1, padding=1, bias=False),

                                       nn.BatchNorm(256),

                                       nn.ReLU(),

                                       nn.Dropout(0.1),

                                       nn.Conv(256, num_classes, kernel_size=1, stride=1, bias=True))

    def execute(self, x, low_level_feat):

        low_level_feat = self.conv1(low_level_feat)

        low_level_feat = self.bn1(low_level_feat)

        low_level_feat = self.relu(low_level_feat)

        x_inter = nn.resize(x, size=(low_level_feat.shape[2], low_level_feat.shape[3]) , mode='bilinear')

        x_concat = concat((x_inter, low_level_feat), dim=1)

        x = self.last_conv(x_concat)

        return x

2.4 完整的模型整合如下： 即将以上部分通过一个类连接起来。

import jittor as jt

from jittor import nn

from jittor import Module

from jittor import init

from jittor.contrib import concat

from decoder import Decoder

from aspp import ASPP

from backbone import resnet50, resnet101

class DeepLab(Module):

    def __init__(self, output_stride=16, num_classes=21):

        super(DeepLab, self).__init__()

        self.backbone = resnet101(output_stride=output_stride)

        self.aspp = ASPP(output_stride)

        self.decoder = Decoder(num_classes)

    def execute(self, input):

        x, low_level_feat = self.backbone(input)

        x = self.aspp(x)

        x = self.decoder(x, low_level_feat)

        x = nn.resize(x, size=(input.shape[2], input.shape[3]), mode='bilinear')

        return x

3. 模型训练

3.1 模型训练参数设定如下：

# Learning parameters

batch_size = 8

learning_rate = 0.005

momentum = 0.9

weight_decay = 1e-4

epochs = 50

3.2 定义模型、优化器、数据加载器。

model = DeepLab(output_stride=16, num_classes=21)

optimizer = nn.SGD(model.parameters(), 

                   lr,

                   momentum=momentum, 

                   weight_decay=weight_decay)

train_loader = TrainDataset(data_root='/vocdata/',

                            split='train',

                            batch_size=batch_size,

                            shuffle=True)

val_loader = ValDataset(data_root='/vocdata/',

                        split='val',

                        batch_size=1,

                        shuffle=False)

3.3 模型训练与验证

# lr scheduler

def poly_lr_scheduler(opt, init_lr, iter, epoch, max_iter, max_epoch):

    new_lr = init_lr * (1 - float(epoch * max_iter + iter) / (max_epoch * max_iter)) ** 0.9

    opt.lr = new_lr

# train function

def train(model, train_loader, optimizer, epoch, init_lr):

    model.train()

    max_iter = len(train_loader)

    for idx, (image, target) in enumerate(train_loader):

        poly_lr_scheduler(optimizer, init_lr, idx, epoch, max_iter, 50) # using poly_lr_scheduler 

        image = image.float32()

        pred = model(image)

        loss = nn.cross_entropy_loss(pred, target, ignore_index=255)

        optimizer.step (loss)

        print ('Training in epoch {} iteration {} loss = {}'.format(epoch, idx, loss.data[0]))

# val function

# we omit evaluator code and you can 

def val (model, val_loader, epoch, evaluator):

    model.eval()

    evaluator.reset()

    for idx, (image, target) in enumerate(val_loader):

        image = image.float32()

        output = model(image)

        pred = output.data

        target = target.data

        pred = np.argmax(pred, axis=1)

        evaluator.add_batch(target, pred)

        print ('Test in epoch {} iteration {}'.format(epoch, idx))

    Acc = evaluator.Pixel_Accuracy()

    Acc_class = evaluator.Pixel_Accuracy_Class()

    mIoU = evaluator.Mean_Intersection_over_Union()

    FWIoU = evaluator.Frequency_Weighted_Intersection_over_Union()

    best_miou = 0.0

    if (mIoU > best_miou):

        best_miou = mIoU

    print ('Testing result of epoch {} miou = {} Acc = {} Acc_class = {} \

                FWIoU = {} Best Miou = {}'.format(epoch, mIoU, Acc, Acc_class, FWIoU, best_miou)) 

3.4 evaluator 写法：使用混淆矩阵计算 Pixel accuracy 和 mIoU。

class Evaluator(object):

    def __init__(self, num_class):

        self.num_class = num_class

        self.confusion_matrix = np.zeros((self.num_class,)*2)

    def Pixel_Accuracy(self):

        Acc = np.diag(self.confusion_matrix).sum() / self.confusion_matrix.sum()

        return Acc

    def Pixel_Accuracy_Class(self):

        Acc = np.diag(self.confusion_matrix) / self.confusion_matrix.sum(axis=1)

        Acc = np.nanmean(Acc)

        return Acc

    def Mean_Intersection_over_Union(self):

        MIoU = np.diag(self.confusion_matrix) / (

                 np.sum(self.confusion_matrix, axis=1) + 

                 np.sum(self.confusion_matrix, axis=0)-

                 np.diag(self.confusion_matrix))

        MIoU = np.nanmean(MIoU)

        return MIoU

    def Frequency_Weighted_Intersection_over_Union(self):

        freq = np.sum(self.confusion_matrix, axis=1) / np.sum(self.confusion_matrix)

        iu = np.diag(self.confusion_matrix) / (

                    np.sum(self.confusion_matrix, axis=1) + np.sum(self.confusion_matrix, axis=0) -

                    np.diag(self.confusion_matrix))

        FWIoU = (freq[freq > 0] * iu[freq > 0]).sum()

        return FWIoU

    def _generate_matrix(self, gt_image, pre_image):

        mask = (gt_image >= 0) & (gt_image < self.num_class)

        label = self.num_class * gt_image[mask].astype('int') + pre_image[mask]

        count = np.bincount(label, minlength=self.num_class**2)

        confusion_matrix = count.reshape(self.num_class, self.num_class)

        return confusion_matrix

    def add_batch(self, gt_image, pre_image):

        assert gt_image.shape == pre_image.shape

        self.confusion_matrix += self._generate_matrix(gt_image, pre_image)

    def reset(self):

        self.confusion_matrix = np.zeros((self.num_class,) * 2)

3.5 训练入口函数

epochs = 50

evaluator = Evaluator(21)

train_loader = TrainDataset(data_root='/voc/data/path/', split='train', batch_size=8, shuffle=True)

val_loader = ValDataset(data_root='/voc/data/path/', split='val', batch_size=1, shuffle=False)

learning_rate = 0.005

momentum = 0.9

weight_decay = 1e-4

optimizer = nn.SGD(model.parameters(), learning_rate, momentum, weight_decay)

for epoch in range (epochs):

    train(model, train_loader, optimizer, epoch, learning_rate)

    val(model, val_loader, epoch, evaluator)

4. 参考

1. pytorch-deeplab-xception
2. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation