Zheng Z, Zheng L, Yang Y. A discriminatively learned CNN embedding for person reidentification[J]. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 2018, 14(1): 13.
从本篇开始复现论文,这里参考的模型结构是著名的IDE网络,这个网络的思想就是结合分类和验证网络来共同学习,使得模型提取得到的行人特征嵌入既能比较两个行人是否是同一人,又能分类得知这个人是谁。而在推理阶段,则是使用分类器之前的特征张量来计算距离。
前期准备
由于模型评估指标计算时间较长,于是采用Cython编写的外部库来计算。如果没安装的同学要注意自行安装,可能需要重新编译。(下载链接 提取码:xerh)
需要使用的库文件如下:
import copy
import os
import random
import re
import time
from collections import defaultdict
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.models as models
import torchvision.transforms as transforms
from PIL import Image
from torch.nn import init
from torch.utils.data import DataLoader, Dataset
from torch.utils.data.sampler import RandomSampler, Sampler
import metrics
try:
from metrics.rank_cylib.rank_cy import evaluate_cy
IS_CYTHON_AVAI = True
except ImportError:
IS_CYTHON_AVAI = False
warnings.warn(
'Cython evaluation (very fast so highly recommended) is '
'unavailable, now use python evaluation.'
)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
首先我们需要定义一个AverageMeter类来帮助我们记录训练过程中的变量:
class AverageMeter(object):
"""Computes and stores the average and current value.
Examples::
>>> # Initialize a meter to record loss
>>> losses = AverageMeter()
>>> # Update meter after every minibatch update
>>> losses.update(loss_value, batch_size)
"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
模型核心代码
首先我们需要定义一个数据集类,将数据集划分为训练集、查询集和测试集:
class Market1501(Dataset):
"""Market1501.
Reference:
Zheng et al. Scalable Person Re-identification: A Benchmark. ICCV 2015.
Dataset statistics:
- identities: 1501 (+1 for background).
- images: 12936 (train) + 3368 (query) + 15913 (gallery).
"""
def __init__(self,root,mode='train',transform=None):
super(Market1501,self).__init__()
self.root = root
self.mode = mode
self.transform = transform
if self.mode == 'train':
self.train_dir = os.path.join(self.root,'bounding_box_train')
self.train = self.process_path(self.train_dir, relabel=True)
self.data = self.train
elif self.mode == 'query':
self.query_dir = os.path.join(self.root,'query')
self.query = self.process_path(self.query_dir, relabel=False)
self.data = self.query
elif self.mode == 'gallery':
self.gallery_dir = os.path.join(self.root,'bounding_box_test')
self.gallery = self.process_path(self.gallery_dir, relabel=False)
self.data = self.gallery
else:
raise ValueError('Invalid mode. Got {}, but expected to be one of [train | query | gallery]'.format(self.mode))
def __len__(self):
return len(self.data)
def __getitem__(self,index):
img_path, pid, camid = self.data[index]
img = Image.open(img_path).convert('RGB')
if self.transform is not None:
img = self.transform(img)
return img, pid, camid, img_path
def process_path(self,path,relabel=False):
img_paths = [os.path.join(path, x) for x in os.listdir(path) if x.endswith('.jpg')]
pattern = re.compile(r'([-\d]+)_c(\d)')
pid_container = set()
for img_path in img_paths:
pid,_ = map(int,pattern.search(img_path).groups())
if pid == -1:
continue
pid_container.add(pid)
pid2label = {pid:label for label, pid in enumerate(pid_container)}
data = []
for img_path in img_paths:
pid,camid = map(int, pattern.search(img_path).groups())
if pid == -1:
continue
camid -= 1
if relabel:
pid = pid2label[pid]
data.append((img_path, pid, camid))
return data
然后就是定义我们的网络结构了,我们这里就使用ResNet-50作为骨干网络。这里需要注意的是我们的输出既要返回分类后的结果用于训练,也要返回特征张量用于推理和评估。这里模型的参数基本都初始化为ResNet-50在ImageNet上预训练的模型,其余参数则是采用了正态分布或者何凯明提出方法进行初始化:
def weights_init_kaiming(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
init.kaiming_normal_(m.weight.data, a=0, mode='fan_in') # For old pytorch, you may use kaiming_normal.
elif classname.find('Linear') != -1:
init.kaiming_normal_(m.weight.data, a=0, mode='fan_out')
init.constant_(m.bias.data, 0.0)
elif classname.find('BatchNorm1d') != -1:
init.normal_(m.weight.data, 1.0, 0.02)
init.constant_(m.bias.data, 0.0)
def weights_init_normal(m):
classname = m.__class__.__name__
if classname.find('Linear') != -1:
init.normal_(m.weight.data, std=0.001)
init.constant_(m.bias.data, 0.0)
class IDE(nn.Module):
def __init__(self, num_classes, reduced_dim=512):
super(IDE, self).__init__()
base_model = models.resnet50(pretrained=True)
base_model.avgpool = nn.AdaptiveAvgPool2d((1,1))
self.model = base_model
add_block = []
add_block += [nn.Linear(2048,reduced_dim)]
add_block += [nn.BatchNorm1d(reduced_dim)]
add_block = nn.Sequential(*add_block)
add_block.apply(weights_init_kaiming)
classifier = []
classifier += [nn.Linear(reduced_dim,num_classes)]
classifier = nn.Sequential(*classifier)
classifier.apply(weights_init_normal)
self.add_block = add_block
self.classifier = classifier
def forward(self, x):
x = self.model.conv1(x)
x = self.model.bn1(x)
x = self.model.relu(x)
x = self.model.maxpool(x)
x = self.model.layer1(x)
x = self.model.layer2(x)
x = self.model.layer3(x)
x = self.model.layer4(x)
x = self.model.avgpool(x)
x = x.view(x.size(0), x.size(1))
# feature
f = self.add_block(x)
# classifier
x = self.classifier(f)
return x, f
训练
加载数据集,需要用到一些数据增强:
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
train_transform = transforms.Compose([
transforms.Resize((256,128)),
transforms.RandomHorizontalFlip(),
transforms.RandomCrop((256,128),padding=(8,16,8,16)),
transforms.ToTensor(),
transforms.Normalize(mean,std)
])
dataset_path = 'D:\\backup\\datasets\\market1501\\Market-1501-v15.09.15' # dataset path
train_dataset = Market1501(dataset_path,mode='train',transform=train_transform)
train_dataloader = DataLoader(
train_dataset,
batch_size=32,
pin_memory=torch.cuda.is_available(),
drop_last=False
)
创建模型,为模型的不同参数设置不同的学习率,并设置损失函数:
model = IDE(751)
lr = 0.0003
ignored_params = list(map(id, model.add_block.parameters()))
ignored_params += list(map(id, model.classifier.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params, model.parameters())
optimizer = torch.optim.Adam([
{'params':base_params,'lr':0.1*lr},
{'params':model.add_block.parameters(),'lr':lr},
{'params':model.classifier.parameters(),'lr':lr}],
lr=lr,
betas=(0.9,0.99),
weight_decay=5e-04
)
criterion = nn.CrossEntropyLoss().to(device)
训练,这里需要默默等待一会儿:
model.to(device)
model.train()
num_batches = len(train_dataloader)
max_epoch = 50
for epoch in range(0, max_epoch):
losses = AverageMeter()
accs = AverageMeter()
batch_time = AverageMeter()
data_time = AverageMeter()
end = time.time()
for batch_idx, data in enumerate(train_dataloader):
data_time.update(time.time() - end)
imgs = data[0]
pids = data[1]
imgs = imgs.to(device)
pids = pids.to(device)
optimizer.zero_grad()
outputs = model(imgs)
preds = outputs[0]
loss = criterion(preds,pids)
loss.backward()
optimizer.step()
batch_time.update(time.time() - end)
losses.update(loss.item(), pids.size(0))
accs.update(metrics.accuracy(preds, pids)[0].item())
if (batch_idx+1) % 10 == 0:
print('Epoch: [{0}/{1}][{2}/{3}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.2f} ({acc.avg:.2f})\t'.format(
epoch+1, max_epoch, batch_idx+1, num_batches,
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accs,
)
)
end = time.time()
if True:
state = {
'state_dict': model.state_dict(),
'epoch': epoch,
'optimizer': optimizer.state_dict()
}
save_path = './model/IDE'
save_filename = os.path.join(save_path,'checkpoint-'+str(epoch)+'.pth.tar')
if not os.path.exists(save_path):
os.makedirs(save_path)
torch.save(state, save_filename)
评估
先把训练好的模型重新加载,同时也把测试数据加载了:
model.to(device)
state = torch.load('model/IDE/checkpoint-49.pth.tar')
model.load_state_dict(state['state_dict'])
test_transform = transforms.Compose([
transforms.Resize((256,128)),
transforms.ToTensor(),
transforms.Normalize(mean,std)
])
query_dataset = Market1501(dataset_path,mode='query',transform=test_transform)
gallery_dataset = Market1501(dataset_path,mode='gallery',transform=test_transform)
query_dataloader = DataLoader(
query_dataset,
batch_size=32,
pin_memory=torch.cuda.is_available(),
drop_last=False
)
gallery_dataloader = DataLoader(
gallery_dataset,
batch_size=32,
pin_memory=torch.cuda.is_available(),
drop_last=False
)
从查询集和测试集提取特征张量:
print('Extracting features from query set ...')
qf, q_pids, q_camids = [], [], []
for batch_idx, data in enumerate(query_dataloader):
imgs, pids, camids = data[0], data[1], data[2]
imgs = imgs.to(device)
model.eval()
features = model(imgs)[1]
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf,0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print('Done, obtained {}-by-{} matrix'.format(qf.size(0), qf.size(1)))
print('Extracting features from gallery set ...')
gf, g_pids, g_camids = [], [], []
for batch_idx, data in enumerate(gallery_dataloader):
imgs, pids, camids = data[0], data[1], data[2]
imgs = imgs.to(device)
model.eval()
features = model(imgs)[1]
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print('Done, obtained {}-by-{} matrix'.format(gf.size(0), gf.size(1)))
计算距离矩阵,这里可以对特征归一化和重排序:
# normalize, better performance
# qf = F.normalize(qf, p=2, dim=1)
# gf = F.normalize(gf, p=2, dim=1)
distmat = metrics.compute_distance_matrix(qf, gf, 'euclidean') # cosine may have better performance
distmat = distmat.numpy()
# re-ranking, better performance
# distmat_qq = metrics.compute_distance_matrix(qf, qf, 'euclidean')
# distmat_gg = metrics.compute_distance_matrix(gf, gf, 'euclidean')
# distmat = re_ranking(distmat, distmat_qq, distmat_gg)
如果需要re-ranking,需要以下函数:
def re_ranking(q_g_dist, q_q_dist, g_g_dist, k1=20, k2=6, lambda_value=0.3):
# The following naming, e.g. gallery_num, is different from outer scope.
# Don't care about it.
original_dist = np.concatenate(
[np.concatenate([q_q_dist, q_g_dist], axis=1),
np.concatenate([q_g_dist.T, g_g_dist], axis=1)],
axis=0)
original_dist = np.power(original_dist, 2).astype(np.float32)
original_dist = np.transpose(1. * original_dist/np.max(original_dist,axis = 0))
V = np.zeros_like(original_dist).astype(np.float32)
initial_rank = np.argsort(original_dist).astype(np.int32)
query_num = q_g_dist.shape[0]
gallery_num = q_g_dist.shape[0] + q_g_dist.shape[1]
all_num = gallery_num
for i in range(all_num):
# k-reciprocal neighbors
forward_k_neigh_index = initial_rank[i,:k1+1]
backward_k_neigh_index = initial_rank[forward_k_neigh_index,:k1+1]
fi = np.where(backward_k_neigh_index==i)[0]
k_reciprocal_index = forward_k_neigh_index[fi]
k_reciprocal_expansion_index = k_reciprocal_index
for j in range(len(k_reciprocal_index)):
candidate = k_reciprocal_index[j]
candidate_forward_k_neigh_index = initial_rank[candidate,:int(np.around(k1/2.))+1]
candidate_backward_k_neigh_index = initial_rank[candidate_forward_k_neigh_index,:int(np.around(k1/2.))+1]
fi_candidate = np.where(candidate_backward_k_neigh_index == candidate)[0]
candidate_k_reciprocal_index = candidate_forward_k_neigh_index[fi_candidate]
if len(np.intersect1d(candidate_k_reciprocal_index,k_reciprocal_index))> 2./3*len(candidate_k_reciprocal_index):
k_reciprocal_expansion_index = np.append(k_reciprocal_expansion_index,candidate_k_reciprocal_index)
k_reciprocal_expansion_index = np.unique(k_reciprocal_expansion_index)
weight = np.exp(-original_dist[i,k_reciprocal_expansion_index])
V[i,k_reciprocal_expansion_index] = 1.*weight/np.sum(weight)
original_dist = original_dist[:query_num,]
if k2 != 1:
V_qe = np.zeros_like(V,dtype=np.float32)
for i in range(all_num):
V_qe[i,:] = np.mean(V[initial_rank[i,:k2],:],axis=0)
V = V_qe
del V_qe
del initial_rank
invIndex = []
for i in range(gallery_num):
invIndex.append(np.where(V[:,i] != 0)[0])
jaccard_dist = np.zeros_like(original_dist,dtype = np.float32)
for i in range(query_num):
temp_min = np.zeros(shape=[1,gallery_num],dtype=np.float32)
indNonZero = np.where(V[i,:] != 0)[0]
indImages = []
indImages = [invIndex[ind] for ind in indNonZero]
for j in range(len(indNonZero)):
temp_min[0,indImages[j]] = temp_min[0,indImages[j]]+ np.minimum(V[i,indNonZero[j]],V[indImages[j],indNonZero[j]])
jaccard_dist[i] = 1-temp_min/(2.-temp_min)
final_dist = jaccard_dist*(1-lambda_value) + original_dist*lambda_value
del original_dist
del V
del jaccard_dist
final_dist = final_dist[:query_num,query_num:]
return final_dist
最后算一下rank-1和mAP:
cmc, mAP = metrics.evaluate_rank(
distmat,
q_pids,
g_pids,
q_camids,
g_camids,
)
print(cmc[0], mAP)
参考
[1] Zhou K, Xiang T. Torchreid: A Library for Deep Learning Person Re-Identification in Pytorch[J]. arXiv preprint arXiv:1910.10093, 2019.