赞
踩
摆了两周,突然觉得不能一直再颓废下去了,应该利用好时间,并且上个月就读了一些经典的图像分割论文比如FCN、UNet和Mask R-CNN,但仅仅只是读了论文并且大概了解了图像分割是在做什么任务的,于是今天就拉动手复现一下,因为只有代码运行起来了,才能进行接下来的代码阅读以及其他改进迁移等后续工作。
本文着重在于代码的复现,其他相关知识会涉及得较少,需要读者自行了解。
看完这篇文章,您将收获一个完整的图像分割项目(一个通用的图像分割数据集及一份可正常执行的代码)。
图来自FCN,Jonathan Long,Evan Shelhamer,Trevor Darrell CVPR2015
图像分割可以大致为实例分割、语义分割,其中语义分割(Semantic Segmentation)是对图像中每一个像素点进行分类,确定每个点的类别(如属于背景、人或车等),从而进行区域划分。目前,语义分割已经被广泛应用于自动驾驶、无人机落点判定等场景中。
FCN全程Fully Convolutional Networks,最早发表于CVPR2015,原论文链接如下:
FCN论文链接:https://arxiv.org/abs/1411.4038
正如其名称全卷积网络,实则是将早年的网络比如VGG的全连接层代替为卷积层,这样做的目的是让模型可以输入不同尺寸的图像,因为全连接层一旦被创建输入输出维度都是固定的,追根溯源就是输入图片的尺寸固定,并且语义分割是像素级别操作,替换为卷积层也更加合理(卷积操作就是像素级别,这些都是后话了)。
更具体的学习视频可以跳转到b站FCN网络结构详解(语义分割)
进入FCN论文链接,点击Code&Data再进入Community Code跳转到paperwithcode网站。
很神奇地是会发现有两个FCN的检索链接,本文所需要的pytorch项目代码在红框这个链接中
Star最高的就是本文所需项目,这个大佬还有自己的个人网页,而且号称是FCN最简单的实现,我可以作证此言不虚,的确是众多代码中最简洁明朗的。
CityScapes数据集官方下载链接:CityScapes Download
然而下载这个数据集需要注册账号,而且需要的是教育邮箱,可能是按照是否带edu.cn域名判断的吧,本人使用学校邮箱成功注册下载了数据集。读者若有不便可以上网其他途径获取或淘宝买个账号。
只需下载前3个数据集即可,gtFine_trainvaltest是精确标注(最主要最关键部分),gtCoarse是粗略标注,leftimg8bit_trainvaltest是原图。虽然模型训练的时候只需要用到gtFine但是因为接下来还需要预处理数据集,因此要将三个数据集下载好,才能执行官方给的预处理代码。
重构数据集
将三个zip解压然后新建一个文件夹命名为CityScapes,然后将三个解压文件里的内容按上图目录放置好,为数据集预处理做准备。
这里需要先下载官方的脚本:cityscapesScripts
接下来对其中的一些地方进行修改,最重要的两个文件为项目下cityscapesscripts\helpers\labels.py和cityscapesscripts\preparation\createTrainIdLabelImgs.py。
蓝色框为原本的代码,直接注释掉添加红框处代码,即指定自己本地的数据集目录,比如我就将CityScapes放到了E盘的dataset目录下。
然后是在label.py文件里按照训练的需要更改trainid,255为不被模型所需要的id,因为FCN中为19类+背景板,所以为20类,刚好符合所以不需要更改label文件中任何内容。
最后运行createTrainIdLabelImgs.py,如果报错的话大概率是因为缺少上图蓝框所示的库,将其直接注释掉就可以了。
之所以需要修改是因为原本的代码里面数据预处理那块太慢了,Cityscapes_utils.py要将trainId写入npy文件,运行速度极慢,这也是先前用官方预处理脚本cityscapesScripts来预处理的原因,预处理的目的其实也只是生成TrainIds的mask图片,和labelIds的png图片是同理的,只是每个像素所对应类别按照label.py里面的label表进行改变。
其实pytorch官方有给出加载CityScapes的数据集代码,但其直接拿来用并不能满足我们要求,所以需要修改一下,就原项目代码的Cityscapes_loader.py和torchvision.datasets.Cityscapes的代码结合,得到如下可执行代码。读者只需用其替换train.py文件即可。
# -*- coding: utf-8 -*- # Author: Reganzhx from __future__ import print_function import random from tqdm import tqdm # 由于训练缓慢,添加进度条方便观察 import imageio import torch import torch.nn as nn import torch.optim as optim from torch.optim import lr_scheduler from torch.autograd import Variable from torch.utils.data import DataLoader from fcn import VGGNet, FCN32s, FCN16s, FCN8s, FCNs # from Cityscapes_loader import CityScapesDataset from CamVid_loader import CamVidDataset from torchvision.datasets import Cityscapes from matplotlib import pyplot as plt import numpy as np import time import sys import os from PIL import Image class CityScapesDataset(Cityscapes): def __init__(self, root: str, split: str = "train", mode: str = "fine", target_type="semantic", transform=None, target_transform=None, transforms=None): super(CityScapesDataset, self).__init__(root, split, mode, target_type, transform, target_transform, transforms) self.means = np.array([103.939, 116.779, 123.68]) / 255. self.n_class = 20 self.new_h = 512 # 数据集图片过大,需要剪裁 self.new_w = 1024 def __getitem__(self, index): img = imageio.imread(self.images[index], pilmode='RGB') targets = [] for i, t in enumerate(self.target_type): if t == "polygon": target = self._load_json(self.targets[index][i]) else: target = imageio.imread(self.targets[index][i]) targets.append(target) target = tuple(targets) if len(targets) > 1 else targets[0] # 针对多目标 可不关注 h, w, _ = img.shape top = random.randint(0, h - self.new_h) left = random.randint(0, w - self.new_w) img = img[top:top + self.new_h, left:left + self.new_w] label = target[top:top + self.new_h, left:left + self.new_w] # reduce mean img = img[:, :, ::-1] # switch to BGR img = np.transpose(img, (2, 0, 1)) / 255. img[0] -= self.means[0] img[1] -= self.means[1] img[2] -= self.means[2] # convert to tensor img = torch.from_numpy(img.copy()).float() label = torch.from_numpy(label.copy()).long() # create one-hot encoding h, w = label.size() target = torch.zeros(self.n_class, h, w) for c in range(self.n_class): target[c][label == c] = 1 sample = {'X': img, 'Y': target, 'l': label} return sample def __len__(self) -> int: return len(self.images) def _get_target_suffix(self, mode: str, target_type: str) -> str: if target_type == "instance": return f"{mode}_instanceIds.png" elif target_type == "semantic": # 让其指向预处理好的target图片 return f"{mode}_labelTrainIds.png" elif target_type == "color": return f"{mode}_color.png" else: return f"{mode}_polygons.json" n_class = 20 batch_size = 2 # 根据测试,1batch需要2G显存,请按实际设置 epochs = 500 lr = 1e-4 momentum = 0 w_decay = 1e-5 step_size = 50 gamma = 0.5 configs = "FCNs-BCEWithLogits_batch{}_epoch{}_RMSprop_scheduler-step{}-gamma{}_lr{}_momentum{}_w_decay{}".format( batch_size, epochs, step_size, gamma, lr, momentum, w_decay) print("Configs:", configs) # create dir for model model_dir = "models" if not os.path.exists(model_dir): os.makedirs(model_dir) model_path = os.path.join(model_dir, configs) use_gpu = torch.cuda.is_available() num_gpu = list(range(torch.cuda.device_count())) # 自行更改root train_data = CityScapesDataset(root='E:/datasets/CityScapes', split='train', mode='fine', target_type='semantic') train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=True) val_data = CityScapesDataset(root='E:/datasets/CityScapes', split='val', mode='fine', target_type='semantic') val_loader = DataLoader(val_data, batch_size=1) vgg_model = VGGNet(requires_grad=True, remove_fc=True) fcn_model = FCNs(pretrained_net=vgg_model, n_class=n_class) if use_gpu: ts = time.time() vgg_model = vgg_model.cuda() fcn_model = fcn_model.cuda() fcn_model = nn.DataParallel(fcn_model, device_ids=num_gpu) print("Finish cuda loading, time elapsed {}".format(time.time() - ts)) criterion = nn.BCEWithLogitsLoss() optimizer = optim.RMSprop(fcn_model.parameters(), lr=lr, momentum=momentum, weight_decay=w_decay) scheduler = lr_scheduler.StepLR(optimizer, step_size=step_size, gamma=gamma) # decay LR by a factor of 0.5 every 30 epochs # create dir for score score_dir = os.path.join("scores", configs) if not os.path.exists(score_dir): os.makedirs(score_dir) IU_scores = np.zeros((epochs, n_class)) pixel_scores = np.zeros(epochs) def train(): for epoch in range(epochs): scheduler.step() ts = time.time() for iter, batch in enumerate(tqdm(train_loader)): optimizer.zero_grad() if use_gpu: inputs = Variable(batch['X'].cuda()) labels = Variable(batch['Y'].cuda()) else: inputs, labels = Variable(batch['X']), Variable(batch['Y']) outputs = fcn_model(inputs) loss = criterion(outputs, labels) loss.backward() optimizer.step() if iter % 10 == 0: print("epoch{}, iter{}, loss: {}".format(epoch, iter, loss.item())) print("Finish epoch {}, time elapsed {}".format(epoch, time.time() - ts)) torch.save(fcn_model, model_path) val(epoch) def val(epoch): fcn_model.eval() total_ious = [] pixel_accs = [] for iter, batch in enumerate(val_loader): if use_gpu: inputs = Variable(batch['X'].cuda()) else: inputs = Variable(batch['X']) output = fcn_model(inputs) output = output.data.cpu().numpy() N, _, h, w = output.shape pred = output.transpose(0, 2, 3, 1).reshape(-1, n_class).argmax(axis=1).reshape(N, h, w) target = batch['l'].cpu().numpy().reshape(N, h, w) for p, t in zip(pred, target): total_ious.append(iou(p, t)) pixel_accs.append(pixel_acc(p, t)) # Calculate average IoU total_ious = np.array(total_ious).T # n_class * val_len ious = np.nanmean(total_ious, axis=1) pixel_accs = np.array(pixel_accs).mean() print("epoch{}, pix_acc: {}, meanIoU: {}, IoUs: {}".format(epoch, pixel_accs, np.nanmean(ious), ious)) IU_scores[epoch] = ious np.save(os.path.join(score_dir, "meanIU"), IU_scores) pixel_scores[epoch] = pixel_accs np.save(os.path.join(score_dir, "meanPixel"), pixel_scores) # borrow functions and modify it from https://github.com/Kaixhin/FCN-semantic-segmentation/blob/master/main.py # Calculates class intersections over unions def iou(pred, target): ious = [] for cls in range(n_class): pred_inds = pred == cls target_inds = target == cls intersection = pred_inds[target_inds].sum() union = pred_inds.sum() + target_inds.sum() - intersection if union == 0: ious.append(float('nan')) # if there is no ground truth, do not include in evaluation else: ious.append(float(intersection) / max(union, 1)) # print("cls", cls, pred_inds.sum(), target_inds.sum(), intersection, float(intersection) / max(union, 1)) return ious def pixel_acc(pred, target): correct = (pred == target).sum() total = (target == target).sum() return correct / total if __name__ == "__main__": val(0) # show the accuracy before training train()
分别在自己办公电脑1030显卡(显存4G)和3060显卡(显存12G)上测试,根据两台电脑运行上看每增加1batch就需要消耗2G显存,因为3060上最大只能将batch size设置为6。3060显卡上1个epoch需要8min,也就是说训练完500epoch需要三天时间,可见图像分割真的是极其消耗资源。而1030上1代竟然耗时2h20min,所以按照时间来看首选设备是3090,这样才可能在一天之内进行完一次完整500epoch训练。
第1轮迭代后pixel accuracy就有75%,目前到第25轮pixel accuracy达到85%,随着epoch数增加,pixel acc也越来越高,希望其最终能突破90%,原论文中可是达到96%pixel准确率。
下图为3060上训练150epoch的结果,每5epoch进行一次val评估。最后使用matplotlib绘制如下曲线,pixel_acc和meanIoU的获取请读者自行额外编写代码获得,此处仅提供绘图代码。
第135epoch取得最高pixel accuracy=0.8766716842651368,meanIoU=0.3268041800950261
from matplotlib import pyplot as plt x=[i for i in range(0,151,5)] #横坐标 # 此处给出我的数据,浮点数都用round函数取到小数点后7位 pix_acc_list=[0.7520696,0.7918097,0.6557526,0.8310604,0.8453417,0.8509236,0.8534471,0.8378322,0.8489639,0.8563263,0.8538324,0.8572157,0.860767,0.8660216,0.8631711,0.8631837,0.8670352,0.8597714,0.8689239,0.8647407,0.8698506,0.8712046,0.8719427,0.8722804,0.8732114,0.871852,0.8714358,0.8766717,0.86854,0.8661136,0.8761132] meanIoU_list=[0.1333057,0.185366,0.1383637,0.2432535,0.2634509,0.2799635,0.2831553,0.2642947,0.2924905,0.3027259,0.3123738,0.2976701,0.3113799,0.3239229,0.3163488,0.3170467,0.3246953,0.3236825,0.3242375,0.3262411,0.3355112,0.3285704,0.3388148,0.328427,0.3378653,0.3385619,0.3358321,0.3268042,0.3297385,0.3347885,0.3379351] plt.figure() plt.plot(x,pix_acc_list,color='blue',label='pixel acc') plt.plot(x,meanIoU_list,color='red',label='meanIoU') plt.xticks(fontsize=16) plt.yticks(fontsize=16) plt.xlabel('Epoch',fontsize=20) plt.ylabel('Score',fontsize=20) plt.legend(fontsize=16) plt.show()
希望您读到这里能有所收获,本文所参考资料也在文末给出,大家可以查阅获取更多知识细节,后续还将不断完善本文内容,敬请期待……
https://bbs.huaweicloud.com/blogs/306716
https://developer.aliyun.com/article/797607
https://www.cnblogs.com/dotman/p/cityscapes_dataset_tips.html
https://zhuanlan.zhihu.com/p/147195575
https://codeantenna.com/a/uD5sJceaS1
https://blog.csdn.net/zz2230633069/article/details/84591532
https://www.zhihu.com/question/276325769/answer/2418207657
https://blog.csdn.net/zz2230633069/article/details/84668984
https://blog.csdn.net/yumaomi/article/details/124847721
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。