行人属性识别，PA100K

1. 前景介绍

      最近在调研行人属性识别，有不少研究行人属性识别的论文，阅读多篇论文之后，自己独立写了一个工程，专门做行人属性识别。

      工程里面的代码主要参考两篇论文，ResNest对应的论文和Bag of Tricks for Image Classification对应的论文。

      主干网络用的是ResNet的最强改进版本：ResNest50，论文详解请看我之前的博客。

      训练技巧主要运用了《Bag of Tricks for Image Classification with Convolutional Neural Networks》里面的训练技巧，详细讲解请参考我的专栏：模型训练技巧。

      训练数据应用的是行人属性数据集PA100K，关于PA100K的介绍请参考我的博客。

2. 训练部分代码讲解

      一共训练10个属性，分别是：性别、年龄、朝向、是否戴帽子、是否戴眼镜、是否有手提包、是否有肩膀包、是否有背包、上衣类型、下衣类型。

    for epoch in range(1, args.e + 1):
        if epoch > args.warm:
            train_scheduler.step(epoch)

        #training procedure
        net.train()

        for batch_index, (images, labels) in enumerate(train_dataloader):
            if epoch <= args.warm:
                warmup_scheduler.step()

            images = images.cuda()
            labels = labels.cuda()

            optimizer.zero_grad()
            predicts = net(images)
     
            loss_gender = cross_entropy(predicts[0], labels[:, 0].long())
            loss_age = cross_entropy(predicts[1], labels[:, 1].long())
            loss_orientation = cross_entropy(predicts[2], labels[:, 2].long())
            loss_hat = cross_entropy(predicts[3], labels[:, 3].long())
            loss_glasses = cross_entropy(predicts[4], labels[:, 4].long())
            loss_handBag = cross_entropy(predicts[5], labels[:, 5].long())
            loss_shoulderBag = cross_entropy(predicts[6], labels[:, 6].long())
            loss_backBag = cross_entropy(predicts[7], labels[:, 7].long())
            loss_upClothing = cross_entropy(predicts[8], labels[:, 8].long())
            loss_downClothing= cross_entropy(predicts[9], labels[:, 9].long())

            loss = loss_gender + loss_age + loss_orientation + loss_hat + loss_glasses + loss_handBag + loss_shoulderBag + loss_backBag + loss_upClothing + loss_downClothing
            loss.backward()
            optimizer.step()

            n_iter = (epoch - 1) * len(train_dataloader) + batch_index + 1

            if batch_index % 10 == 0:
                print('Training Epoch: {epoch} [{trained_samples}/{total_samples}]\tLoss: {:0.4f}\tLoss_gender: {:0.4f}\tLoss_age: {:0.4f}\tLoss_ori: {:0.4f}\tLoss_hat: {:0.4f}\tLoss_glasses: {:0.4f}\tLoss_handBag: {:0.4f}\t'.format(
                    loss.item(),
                    loss_gender.item(),
                    loss_age.item(),
                    loss_orientation.item(),
                    loss_hat.item(),
                    loss_glasses.item(),
                    loss_handBag.item(),
                    epoch=epoch,
                    trained_samples=batch_index * args.b + len(images),
                    total_samples=len(train_dataloader.dataset),
                ))

            #visualization
            visualize_lastlayer(writer, net, n_iter)
            visualize_train_loss(writer, loss.item(), n_iter)


        visualize_learning_rate(writer, optimizer.param_groups[0]['lr'], epoch)
        visualize_param_hist(writer, net, epoch) 

        net.eval()

        total_loss = 0
        correct = np.zeros(10)
        ignore = np.zeros(10)
        print("=>test model")
        for images, labels in tqdm(test_dataloader):
            images = images.cuda()
            labels = labels.cuda()

            predicts = net(images)

            for index in range(10):
                _, preds = predicts[index].max(1)
                ignore[index] += int((labels[:, index]==-1).sum())
                correct[index] += preds.eq(labels[:, index]).sum().float()

                loss = cross_entropy(predicts[index], labels[:, index].long())
                total_loss += loss.item()

        test_loss = total_loss / len(test_dataloader)
        all_list = np.array([len(test_dataloader.dataset) for i in range(10)])-ignore
        acc_list = correct / all_list
        print(acc_list.tolist())

        训练过程如下：

3. 测试代码讲解

"""
author:guopei
"""
import os
import cv2
import torch
from torch.nn import DataParallel
import numpy as np
from PIL import Image,ImageDraw,ImageFont
import transforms
from models import resnest50


class Person_Attribute(object):
    def __init__(self, weights="resnest50.pth"):
        self.device = torch.device("cuda")
        self.net = resnest50().to(self.device)
        self.net = DataParallel(self.net)
        self.weights = weights
        self.net.load_state_dict(torch.load(self.weights))

        TRAIN_MEAN = [0.485, 0.499, 0.432]
        TRAIN_STD = [0.232, 0.227, 0.266]
        self.transforms = transforms.Compose([
                    transforms.ToCVImage(),
                    transforms.Resize((128,256)),
                    transforms.ToTensor(),
                    transforms.Normalize(TRAIN_MEAN, TRAIN_STD)
        ])

    def recog(self, img_path):
        img = cv2.imread(img_path)
        img = self.transforms(img)
        img = img.unsqueeze(0)

        with torch.no_grad():
            self.net.eval()
            img_input = img.to(self.device)
            outputs = self.net(img_input)
            results = []
            for output in outputs:
                output = torch.softmax(output, 1)
                output = np.array(output[0].cpu())
                label = np.argmax(output)
                score = output[label]
                results.append((label, score))
        return results


name_dict = {
                "gender":['男', '女'],
                "age":["老年", "中年", "少年"],
                "orientation":['正面', '侧面', '背面'],
                "hat":["不戴帽子", "戴帽子"],
                "glasses":["不戴眼镜", "戴眼镜"],
                "handBag":["没有", "有"],
                "shoulderBag":["没有", "有"],
                "backBag":["没有", "有"],
                "upClothing":["短袖", "长袖"],
                "downClothing":["长裤", "短裤", "裙子"]
            }

if __name__ == "__main__":
    atts = ["gender","age", "orientation", "hat", "glasses",
            "handBag", "shoulderBag", "backBag", "upClothing", "downClothing"]

    person_attribute = Person_Attribute("./resnest50-50-regular.pth")
    img_path = "test.jpg"
    results = person_attribute.recog(img_path)
    print(results)
    img = cv2.imread(img_path)
    img = cv2.resize(img, (128,256))
    img1 = img*0 +255
    img1[:,:,0] *= 255
    img1[:,:,2] *= 255

    line = ""
    labels = [i[0] for i in results]
    for att, label in zip(atts, labels):
        if label == -1:
            continue
        line += "%s:%s\n" % (att, name_dict[att][label])


    img1 = cv2.cvtColor(img1, cv2.COLOR_BGR2RGB)
    img1 = Image.fromarray(img1)
    draw = ImageDraw.Draw(img1)
    font = ImageFont.truetype("simhei.ttf", 12, encoding="utf-8")
    draw.text((0, 0), line, (255, 0, 0), font=font)
    img1 = cv2.cvtColor(np.array(img1), cv2.COLOR_RGB2BGR)

    img_rst = np.hstack([img, img1])
    cv2.imwrite(os.path.basename(img_path), img_rst)

          测试效果如下：

 

注：本文所用的训练数据都是公开数据集，如果你需要训练好的模型或者训练数据或者技术交流，请留言或者微信。近期会将代码开源。

注：很多粉丝拿到了训练的源码，对其中train.txt和val.txt有疑问，这里统一回复一下。

train.txt和val.txt的格式如下，每类标签是从0开始的，不确定的类别用-1表示，标签之间用空格分开。

/work/img/1.jpg 0 1 1 1 0 0 0 0 1 1
/work/img/2.jpg 0 1 1 -1 0 0 0 0 1 1
/work/img/3.jpg 0 1 1 0 0 0 0 0 1 1
/work/img/4.jpg 0 1 1 1 0 1 0 0 1 1
/work/img/5.jpg 0 1 1 1 0 1 0 0 1 1
/work/img/6.jpg 0 1 1 1 0 0 0 0 1 1
/work/img/7.jpg 0 1 1 1 0 0 0 0 1 1
/work/img/8.jpg 0 1 1 1 0 0 0 0 1 1
/work/img/9.jpg 0 1 1 1 0 1 0 0 1 1