:red_car: The 1st Place Submission to AICity Challenge 2020 re-id track (Baidu-UTS submission)
License: MIT License
Contact Me:
✉ Email: zhedongzheng AT um.edu.mo
✧ Website: http://zdzheng.xyz
✧ Linkedin: https://www.linkedin.com/in/zhedongzheng/
✧ Google Scholar: https://scholar.google.com/citations?hl=en&user=XT17oUEAAAAJ
hi,
I didn't know the meaning of gallery, and I didn't find any clue about "gallery" in AICity dataset.
thanks
Hi,
It seems that the google drive URL does not exist anymore and the OneDrive URL requires authentication.
Can you take a look and reupload those models somewhere?
Thanks a lot
Hi, I don't have resources to train it. It will be helpful if you could share the trained model.
AICIty-reID-2020/pytorch/fast_submit_baidu.py
Lines 321 to 356 in 6a8e5da
请问郑博士，这一部分代码是在做什么？谢谢！
郑博士您好，在fast_submit.py 中将query_feature0~query_feature12合并作为最终的特征。请问只有feature0是pytorch这边的吗？
没有看到从哪里输入AI city 2020的数据开始处理的，全程没有注释，请问是那个参数是输入数据？
I want to know how much performance has been improved under same circumstances.
I was going through your research paper and was parallelly going through the codes. I was not able to find how you used UNIT(cycle gan) to perform style transfer and then content manipulation.
Kindly help us understand the flow of your code or provide a link where you have trained synthetic images on cycleGAN.
Hello,
Is implemented in the code of this repo the Copy&Paste technique? In that case , in which file is it implemented?
thank you in advance
@layumi 郑博士 我直接运行python train_2020.py能真确开始训练，但是按照README.md上提供的那一长传命令，就会报下面的错，这是为何？
requests.exceptions.SSLError: HTTPSConnectionPool(host='data.lip6.fr', port=443): Max retries exceeded with url: /cadene/pretrainedmodels/se_resnext101_32x4d-3b2fe3d8.pth (Caused by SSLError(SSLError("bad handshake: Error([('SSL routines', 'tls_process_server_certificate', 'certificate verify failed')],)",),))
测试的时候的数据集不应该是所有真实图片的数量，应该是56277 images呀以及"image_query/". This dir contains 1052 images as queries.
可是结果却显示
Gallery Size: 1477400
Query Size: 465
FileNotFoundError: [Errno 2] No such file or directory: './data/pytorch2020/train+virtual'
Hello, run the train_2020.py file as you said, but the above error occurs. What does the train+virtual folder do?
Hello, Dr. Zheng.
Thank you for your excellent work. I have a question for you. In vehicle reidentification, if the three-dimensional point cloud and two-dimensional image are used together, can the accuracy rate of vehicle recognition be improved？
Thank you very much!
Hello,
Can you help me solve this issue. I am testing the pretrained weights and this is the error i am getting
郑博士，您好。 我在尝试基于您提供的Res50_imbalance_s1_256_p0.5_lr2_mt_d0_b48 (TMM) 这个模型进行stage_two 的训练，也就是基于 VeRi 的 fine_tune. 希望可以最终得到和您提供的 ft_Res50_imbalance_s1_256_p0.5_lr1_mt_d0.2_b48_w5 (TMM) 近似mAP的模型。
但是，训练过程中我发现，VeRi数据集中所含车辆的编号大于700，但ft_Res50_imbalance_s1_256_p0.5_lr1_mt_d0.2_b48_w5 (TMM) 最后一层的classifier的输出是576.
您是应该用了一部分VeRi来进行fine-tune的么？您是怎样划分的呢？
如下是我应用nclass=769(VeRi 训练集里所有车辆进行训练），stage_two= 12 后，进行测试得到的结果：
Rank@1:0.846246 Rank@5:0.937425 Rank@10:0.967819 mAP:0.547631
跟您提供的模型的mAP=83.4%相比还差了不少。
郑博士您好，麻烦请教一下在python train_2020.py...的时候出现 “FileNotFoundError: [Errno 2] No such file or directory: './train.py'”，在pytorch这个文件夹下没有train.py这个文件，请问这是什么原因？
hello,where is efficientnet_pytorch?Thanks
请问最终用的哪12个模型进行集成的呀？
I was looking around your research, I try to run the file train_2020.py but it says that it can't import pretrainedmodels
What should I do?
I have read your paper and code roughly, but have few questions.
Your code used SE-ResNeXt101, but ResNeXt101 in your paper; your code did not use Block3; your learning rate is 0.02 instead 0.01 in your paper; the lr scheduler is MultiStepLR instead Cosine; the feature dim is 4092 instead 1024; May I ask why?
I am confused how to provide cropped_aicity folder.
Should I make cropped images folder structure as we do with test_data?
my test_data structure is: test_data/gallery/im_test and test_data/query/im_query
When I run fast_submit_baidu.py I am getting the following error:
RuntimeError: CUDA out of memory. Tried to allocate 1.39 GiB (GPU 0; 7.80 GiB total capacity; 5.19 GiB already allocated; 174.31 MiB free; 5.59 GiB reserved in total by PyTorch)
@layumi Could you please give me some guidance about how to prepare test_dir and crop_dir to run fast_submit_baidu.py?
Thanks again for your great job.
你好，郑博士，我把prepare_2020.py训练好的图像数据放在了如图中1位置红色框的文件夹下，但是运行train_2020.py后，报错在./data/pytorch2020/train路径下找不到任何类，这是怎么回事呢？郑博士，是否可以在readme里写清代码里所有文件路径都是什么？需要把什么样的数据放在对应的哪些文件路径下？
Hello @layumi
I didn't quite understand your mAP calculation.
Is there any explanation you could share me?
Why are you adding olp_precision even though every ap is being summed up
And also I didn't get why we should divide by 2 before sum them up (here: ap = ap + d_recall*(old_precision + precision)/2)
def compute_mAP(index, good_index, junk_index):
ap = 0
cmc = torch.IntTensor(len(index)).zero_()
if good_index.size==0: # if empty
cmc[0] = -1
return ap,cmc
# remove junk_index
mask = np.in1d(index, junk_index, invert=True)
index = index[mask]
# find good_index index
ngood = len(good_index)
mask = np.in1d(index, good_index)
rows_good = np.argwhere(mask==True)
rows_good = rows_good.flatten()
cmc[rows_good[0]:] = 1
for i in range(ngood):
d_recall = 1.0/ngood
precision = (i+1)*1.0/(rows_good[i]+1)
if rows_good[i]!=0:
old_precision = i*1.0/rows_good[i]
else:
old_precision=1.0
ap = ap + d_recall*(old_precision + precision)/2
return ap, cmc请问郑博士，论文中的negative mining对应哪一部分代码？谢谢
作者您好，我是一个新手，我想直接运行测试代码，跳过训练的部分。我尝试运行python fast_submit.py但是发现dataset文件结构不对：
No such file or directory: './data/test_data/gallery'
我参照https://github.com/PaddlePaddle/Research/tree/master/CV/PaddleReid/process_aicity_data 完成了数据准备，但是'./data/test_data/gallery'依然没有，请问测试数据集应该怎么准备？感谢
你好，想问一下数据怎么获取呢？我看官网数据获取需要密码。你那边可以共享一下吗？
你好，在prepare_2020.py中生成了gallery这个文件夹，但并没有程序往这个文件夹中添加，但在运行test_2020.py中会报错，说这个文件夹中没有图片
Thx for you job! Because the data we can't get it, so can you share your pre-trained model for testing? If not but why?
Could you please provide all the necessay features in order to run the fast_submit_baidu.py?
For example the "query_fea_ResNeXt101_vd_64x4d_cos_alldata_final.pkl" is missing
请问训练和测试用的数据集是AICity20 官网的Track2吗？我在官网下载了压缩包，但解压后不是以下格式：
|- 2020AICITY
|- ...
|- 000345_c020_9.jpg
|- ...
|- 002028_c036_4_9_95_2.jpg
请问需要重新组织官网数据集吗？
Hello @layumi @miraclebiu. Thank you for work once more.
I have a question about training camera-aware model.
There is nothing said about in the README. Could you please give me some guidance if possible?
I would like to know how to train and how to implement it.
Thank you a lot
I am running fast_submit_baidu.py file. I have set --batchsize 1. But i am getting an error.
python fast_submit_baidu.py --gpu_ids 0,1 --batchsize 1
./data/test_data
torch.Size([1052, 9718])
Gallery Cluster Class Number: 798
Gallery Cluster Image per Class: 22.92
Low Qualtiy Image in Query: 92
Low Qualtiy Image in Gallery: 2760
torch.Size([1052, 2048])
RuntimeError: CUDA out of memory. Tried to allocate 1.39 GiB (GPU 0; 7.80 GiB total capacity; 5.19 GiB already allocated; 113.00 MiB free; 5.59 GiB reserved in total by PyTorch)
It says `CUDA out of memory. Could you please anyone give me gaudiness?! How can i solve this problem?
郑博士您好！
您在论文中提到有minibatch中５０％的图片用作Negative mining ，并且使用了度量学习的方法．请问Negative mining 以及 triplet loss对应于代码中的何处？
Hi，Professor Zheng.I'm intertested of the Content Manipulation operation to generate a new car.I had read your annother paper《DG-Net》，and run the demo of your code.But I want to use it on the vehicle dataset,I want to know how to realize the opration the same the Content Manipulation.Look forward to your apply,Thanks!
@layumi 这博士你好。evaluate_gpu.py中的第53行， cmc[rows_good[0]:] = 1 这句代码，我感觉不太合适。cmc的长度是36935，是去除junk_index之前的index的长度；而rows_good所代表的索引是去除junk_index之后的index的，此时的index，在经过第45行的“index = index[mask]”之后，已经变短了，为36875。那“ cmc[rows_good[0]:] = 1”这句代码是不是不太合适啊？？
请问如何实现在论文中最后提到的可视化显示
hello，where I can download aicity 2020 data？you have Baidu cloud link？ thank you.
您好，在４６－４７行
parser.add_argument('--test_dir', default='../data/test_data', type=str, help='./test_data')
parser.add_argument('--crop_dir', default='../data/cropped_aicity', type=str, help='./test_data')
并没有发现相关代码
Hi @layumi
sorry for bothering you. May I ask you again?
def extract_feature(/*...*/):
// ...
for data in tqdm(dataloaders):
//...
ff = torch.FloatTensor(n,512).zero_().cuda()
Where did 512 come from ?
I am curious about the reason why you set the parameter as 512.
郑博士以及各位大神好，代码出现点问题，我一个ＥＰＯＣＨ都没跑玩，准确率为１，损失基本为０：
条件是：--name SE_imbalance_s1_384_p0.5_lr2_mt_d0_b24+v+aug --warm_epoch 5 --droprate 0 --stride 1 --erasing_p 0.5 --autoaug --inputsize 384 --lr 0.02 --use_SE --gpu_ids 0 --train_virtual --batchsize 8
from future import print_function, division
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable
from torchvision import datasets, transforms
import torch.backends.cudnn as cudnn
import matplotlib
matplotlib.use('agg')
import matplotlib.pyplot as plt
import time
import os
from losses import AngleLoss, ArcLoss
from model import ft_net, ft_net_dense, ft_net_EF4, ft_net_EF5, ft_net_EF6, ft_net_IR, ft_net_NAS, ft_net_SE,
ft_net_DSE, PCB, CPB, ft_net_angle, ft_net_arc
from random_erasing import RandomErasing
import yaml
from AugFolder import AugFolder
from shutil import copyfile
import random
from autoaugment import ImageNetPolicy
from utils import get_model_list, load_network, save_network, make_weights_for_balanced_classes
version = torch.version
try:
from apex.fp16_utils import *
from apex import amp, optimizers
except ImportError: # will be 3.x series
print(
'This is not an error. If you want to use low precision, i.e., fp16, please install the apex with cuda support (https://github.com/NVIDIA/apex) and update pytorch to 1.0')
if not os.path.isdir('/home/ubuntu-guangzhaodai/Desktop/AICIty-reID-2020/data/outputs'):
os.mkdir('/home/ubuntu-guangzhaodai/Desktop/AICIty-reID-2020/data/outputs')
######################################################################
parser = argparse.ArgumentParser(description='Training')
parser.add_argument('--gpu_ids', default='0', type=str, help='gpu_ids: e.g. 0 0,1,2 0,2')
parser.add_argument('--adam', action='store_true', help='use all training data')
parser.add_argument('--name', default='ft_ResNet50', type=str, help='output model name')
parser.add_argument('--init_name', default='imagenet', type=str, help='initial with ImageNet')
parser.add_argument('--data_dir', default='/home/ubuntu-guangzhaodai/Desktop/AICIty-reID-2020/data/pytorch2020',
type=str, help='training dir path')
parser.add_argument('--train_all', action='store_true', help='use all training data')
parser.add_argument('--train_veri', action='store_true', help='use part training data + veri')
parser.add_argument('--train_virtual', action='store_true', help='use part training data + virtual')
parser.add_argument('--train_comp', action='store_true', help='use part training data + comp')
parser.add_argument('--train_pku', action='store_true', help='use part training data + pku')
parser.add_argument('--train_comp_veri', action='store_true', help='use part training data + comp +veri')
parser.add_argument('--train_milktea', action='store_true', help='use part training data + com + veri+pku')
parser.add_argument('--color_jitter', action='store_true', help='use color jitter in training')
parser.add_argument('--batchsize', default=32, type=int, help='batchsize')
parser.add_argument('--inputsize', default=299, type=int, help='batchsize')
parser.add_argument('--h', default=299, type=int, help='height')
parser.add_argument('--w', default=299, type=int, help='width')
parser.add_argument('--stride', default=2, type=int, help='stride')
parser.add_argument('--pool', default='avg', type=str, help='last pool')
parser.add_argument('--autoaug', action='store_true', help='use Color Data Augmentation')
parser.add_argument('--erasing_p', default=0, type=float, help='Random Erasing probability, in [0,1]')
parser.add_argument('--use_dense', action='store_true', help='use densenet121')
parser.add_argument('--use_NAS', action='store_true', help='use nasnetalarge')
parser.add_argument('--use_SE', action='store_true', help='use se_resnext101_32x4d')
parser.add_argument('--use_DSE', action='store_true', help='use senet154')
parser.add_argument('--use_IR', action='store_true', help='use InceptionResNetv2')
parser.add_argument('--use_EF4', action='store_true', help='use EF4')
parser.add_argument('--use_EF5', action='store_true', help='use EF5')
parser.add_argument('--use_EF6', action='store_true', help='use EF6')
parser.add_argument('--lr', default=0.05, type=float, help='learning rate')
parser.add_argument('--droprate', default=0.5, type=float, help='drop rate')
parser.add_argument('--PCB', action='store_true', help='use PCB+ResNet50')
parser.add_argument('--CPB', action='store_true', help='use Center+ResNet50')
parser.add_argument('--fp16', action='store_true',
help='use float16 instead of float32, which will save about 50% memory')
parser.add_argument('--balance', action='store_true', help='balance sample')
parser.add_argument('--angle', action='store_true', help='use angle loss')
parser.add_argument('--arc', action='store_true', help='use arc loss')
parser.add_argument('--warm_epoch', default=0, type=int, help='the first K epoch that needs warm up')
parser.add_argument('--resume', action='store_true', help='use arc loss')
opt = parser.parse_args()
if opt.resume:
model, opt, start_epoch = load_network(opt.name, opt)
else:
start_epoch = 0
print(start_epoch)
fp16 = opt.fp16
data_dir = opt.data_dir
name = opt.name
if not opt.resume:
str_ids = opt.gpu_ids.split(',')
gpu_ids = []
for str_id in str_ids:
gid = int(str_id)
if gid >= 0:
gpu_ids.append(gid)
opt.gpu_ids = gpu_ids
if len(opt.gpu_ids) > 0:
cudnn.enabled = True
cudnn.benchmark = True
######################################################################
if opt.h == opt.w:
transform_train_list = [
# transforms.RandomRotation(30),
transforms.Resize((opt.inputsize, opt.inputsize), interpolation=3),
transforms.Pad(15),
# transforms.RandomCrop((256,256)),
transforms.RandomResizedCrop(size=opt.inputsize, scale=(0.75, 1.0), ratio=(0.75, 1.3333), interpolation=3),
# Image.BICUBIC)
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
transform_val_list = [
transforms.Resize(size=opt.inputsize, interpolation=3), # Image.BICUBIC
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
else:
transform_train_list = [
# transforms.RandomRotation(30),
transforms.Resize((opt.h, opt.w), interpolation=3),
transforms.Pad(15),
# transforms.RandomCrop((256,256)),
transforms.RandomResizedCrop(size=(opt.h, opt.w), scale=(0.75, 1.0), ratio=(0.75, 1.3333), interpolation=3),
# Image.BICUBIC)
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
transform_val_list = [
transforms.Resize((opt.h, opt.w), interpolation=3), # Image.BICUBIC
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
if opt.PCB:
transform_train_list = [
transforms.Resize((384, 192), interpolation=3),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
transform_val_list = [
transforms.Resize(size=(384, 192), interpolation=3), # Image.BICUBIC
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
if opt.erasing_p > 0:
transform_train_list = transform_train_list + [RandomErasing(probability=opt.erasing_p, mean=[0.0, 0.0, 0.0])]
if opt.color_jitter:
transform_train_list = [transforms.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1,
hue=0)] + transform_train_list
transform_train_list_aug = [ImageNetPolicy()] + transform_train_list
print(transform_train_list)
data_transforms = {
'train': transforms.Compose(transform_train_list),
'train_aug': transforms.Compose(transform_train_list_aug),
'val': transforms.Compose(transform_val_list),
}
train_all = ''
if opt.train_all:
train_all = '_all'
if opt.train_veri:
train_all = '+veri'
if opt.train_comp:
train_all = '+comp'
if opt.train_virtual:
train_all = '+virtual'
if opt.train_pku:
train_all = '+pku'
if opt.train_comp_veri:
train_all = '+comp+veri'
if opt.train_milktea:
train_all = '+comp+veri+pku'
image_datasets = {}
if not opt.autoaug:
image_datasets['train'] = datasets.ImageFolder(os.path.join(data_dir, 'train' + train_all),
data_transforms['train'])
else:
image_datasets['train'] = AugFolder(os.path.join(data_dir, 'train' + train_all),
data_transforms['train'], data_transforms['train_aug'])
if opt.balance:
dataset_train = image_datasets['train']
weights = make_weights_for_balanced_classes(dataset_train.imgs, len(dataset_train.classes))
weights = torch.DoubleTensor(weights)
sampler = torch.utils.data.sampler.WeightedRandomSampler(weights, len(weights))
dataloaders = {}
dataloaders['train'] = torch.utils.data.DataLoader(image_datasets['train'], batch_size=opt.batchsize,
sampler=sampler, num_workers=8,
pin_memory=True) # 8 workers may work faster
else:
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=opt.batchsize,
shuffle=True, num_workers=8, pin_memory=True)
# 8 workers may work faster
for x in ['train']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train']}
class_names = image_datasets['train'].classes
use_gpu = torch.cuda.is_available()
######################################################################
scheduler is an LR scheduler object fromtorch.optim.lr_scheduler.y_loss = {} # loss history
y_loss['train'] = []
y_loss['val'] = []
y_err = {}
y_err['train'] = []
y_err['val'] = []
def train_model(model, criterion, optimizer, scheduler, start_epoch=0, num_epochs=25):
since = time.time()
warm_up = 0.1 # We start from the 0.1*lrRate
gamma = 0.0 # auto_aug
warm_iteration = round(dataset_sizes['train'] / opt.batchsize) * opt.warm_epoch # first 5 epoch
total_iteration = round(dataset_sizes['train'] / opt.batchsize) * num_epochs
best_model_wts = model.state_dict()
best_loss = 9999
best_epoch = 0
for epoch in range(num_epochs - start_epoch):
epoch = epoch + start_epoch
print('gamma: %.4f' % gamma)
print('Epoch {}/{}'.format(epoch, num_epochs - 1))
print('-' * 50)
# Each epoch has a training and validation phase
for phase in ['train']:
if phase == 'train':
scheduler.step()
model.train(True) # Set model to training mode
else:
model.train(False) # Set model to evaluate mode
running_loss = 0.0
running_corrects = 0.0
global_step = 0
iterate_num = 0
# Iterate over data.
for data in dataloaders[phase]:
# get the inputs
if opt.autoaug:
inputs, inputs2, labels = data
if random.uniform(0, 1) > gamma:
inputs = inputs2
gamma = min(1.0, gamma + 1.0 / total_iteration)
else:
inputs, labels = data
now_batch_size, c, h, w = inputs.shape
if now_batch_size < opt.batchsize: # skip the last batch
continue
# print(inputs.shape)
# wrap them in Variable
if use_gpu:
inputs = Variable(inputs.cuda().detach())
labels = Variable(labels.cuda().detach())
else:
inputs, labels = Variable(inputs), Variable(labels)
# if we use low precision, input also need to be fp16
# if fp16:
# inputs = inputs.half()
# zero the parameter gradients
optimizer.zero_grad()
# forward
if phase == 'val':
with torch.no_grad():
outputs = model(inputs)
else:
outputs = model(inputs)
if opt.PCB:
part = {}
sm = nn.Softmax(dim=1)
num_part = 6
for i in range(num_part):
part[i] = outputs[i]
score = sm(part[0]) + sm(part[1]) + sm(part[2]) + sm(part[3]) + sm(part[4]) + sm(part[5])
_, preds = torch.max(score.data, 1)
loss = criterion(part[0], labels)
for i in range(num_part - 1):
loss += criterion(part[i + 1], labels)
elif opt.CPB:
part = {}
sm = nn.Softmax(dim=1)
num_part = 4
for i in range(num_part):
part[i] = outputs[i]
score = sm(part[0]) + sm(part[1]) + sm(part[2]) + sm(part[3])
_, preds = torch.max(score.data, 1)
loss = criterion(part[0], labels)
for i in range(num_part - 1):
loss += criterion(part[i + 1], labels)
else:
loss = criterion(outputs, labels)
if opt.angle or opt.arc:
outputs = outputs[0]
_, preds = torch.max(outputs.data, 1)
# backward + optimize only if in training phase
if epoch < opt.warm_epoch and phase == 'train':
warm_up = min(1.0, warm_up + 0.9 / warm_iteration)
loss *= warm_up
# backward + optimize only if in training phase
if phase == 'train':
if fp16: # we use optimier to backward loss
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
global_step += 1
iterate_num += now_batch_size
print('Epoch: [{0}] [{1} / {2}]\t'
'Global_step: {3:}\t'
'Loss: {4:.3f}\t'
'Accurcy: {5:.3f}\t'.format(epoch, iterate_num, dataset_sizes[phase], global_step, loss.item(),
float(torch.sum(preds == labels.data)) / now_batch_size))
# print('Epoch:%d Iteration:%d Total:%d Global_step:%d loss:%.2f accuracy:%.2f' % (
# epoch, iterate_num, dataset_sizes[phase], global_step, loss.item(), float(torch.sum(preds == labels.data)) / now_batch_size))
# statistics
if int(version[0]) > 0 or int(version[2]) > 3: # for the new version like 0.4.0, 0.5.0 and 1.0.0
running_loss += loss.item() * now_batch_size
else: # for the old version like 0.3.0 and 0.3.1
running_loss += loss.data[0] * now_batch_size
running_corrects += float(torch.sum(preds == labels.data))
del (loss, outputs, inputs, preds)
epoch_loss = running_loss / dataset_sizes[phase]
epoch_acc = running_corrects / dataset_sizes[phase]
print('{} Loss: {:.4f} Acc: {:.4f}'.format(
phase, epoch_loss, epoch_acc))
y_loss[phase].append(epoch_loss)
y_err[phase].append(1.0 - epoch_acc)
# deep copy the model
if len(opt.gpu_ids) > 1:
save_network(model.module, opt.name, epoch + 1)
else:
save_network(model, opt.name, epoch + 1)
draw_curve(epoch)
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print()
if epoch_loss < best_loss:
best_loss = epoch_loss
best_epoch = epoch
last_model_wts = model.state_dict()
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print('Best epoch: {:d} Best Train Loss: {:4f}'.format(best_epoch, best_loss))
# load best model weights
model.load_state_dict(last_model_wts)
save_network(model, opt.name, 'last')
return model
######################################################################
x_epoch = []
fig = plt.figure()
ax0 = fig.add_subplot(121, title="loss")
ax1 = fig.add_subplot(122, title="top1err")
def draw_curve(current_epoch):
x_epoch.append(current_epoch)
ax0.plot(x_epoch, y_loss['train'], 'bo-', label='train')
# ax0.plot(x_epoch, y_loss['val'], 'ro-', label='val')
ax1.plot(x_epoch, y_err['train'], 'bo-', label='train')
# ax1.plot(x_epoch, y_err['val'], 'ro-', label='val')
if current_epoch == 0:
ax0.legend()
ax1.legend()
fig.savefig(os.path.join('/home/ubuntu-guangzhaodai/Desktop/AICIty-reID-2020/data/outputs', name, 'train.png'))
######################################################################
if not opt.resume:
opt.nclasses = len(class_names)
if opt.use_dense:
model = ft_net_dense(len(class_names), opt.droprate, opt.stride, None, opt.pool)
elif opt.use_NAS:
model = ft_net_NAS(len(class_names), opt.droprate, opt.stride)
elif opt.use_SE:
model = ft_net_SE(len(class_names), opt.droprate, opt.stride, opt.pool)
elif opt.use_DSE:
model = ft_net_DSE(len(class_names), opt.droprate, opt.stride, opt.pool)
elif opt.use_IR:
model = ft_net_IR(len(class_names), opt.droprate, opt.stride)
elif opt.use_EF4:
model = ft_net_EF4(len(class_names), opt.droprate)
elif opt.use_EF5:
model = ft_net_EF5(len(class_names), opt.droprate)
elif opt.use_EF6:
model = ft_net_EF6(len(class_names), opt.droprate)
else:
model = ft_net(len(class_names), opt.droprate, opt.stride, None, opt.pool)
if opt.PCB:
model = PCB(len(class_names))
if opt.CPB:
model = CPB(len(class_names))
if opt.angle:
model = ft_net_angle(len(class_names), opt.droprate, opt.stride)
elif opt.arc:
model = ft_net_arc(len(class_names), opt.droprate, opt.stride)
if opt.init_name != 'imagenet':
old_opt = parser.parse_args()
init_model, old_opt, _ = load_network(opt.init_name, old_opt)
print(old_opt)
opt.stride = old_opt.stride
opt.pool = old_opt.pool
opt.use_dense = old_opt.use_dense
if opt.use_dense:
model = ft_net_dense(opt.nclasses, droprate=opt.droprate, stride=opt.stride, init_model=init_model,
pool=opt.pool)
else:
model = ft_net(opt.nclasses, droprate=opt.droprate, stride=opt.stride, init_model=init_model, pool=opt.pool)
##########################
if start_epoch >= 60:
opt.lr = opt.lr * 0.1
if start_epoch >= 75:
opt.lr = opt.lr * 0.1
if len(opt.gpu_ids) > 1:
model = torch.nn.DataParallel(model, device_ids=opt.gpu_ids).cuda()
if not opt.CPB:
ignored_params = list(map(id, model.module.classifier.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params, model.parameters())
optimizer_ft = optim.SGD([
{'params': base_params, 'lr': 0.1 * opt.lr},
{'params': model.module.classifier.parameters(), 'lr': opt.lr}
], weight_decay=5e-4, momentum=0.9, nesterov=True)
else:
ignored_params = (list(map(id, model.module.classifier0.parameters()))
+ list(map(id, model.module.classifier1.parameters()))
+ list(map(id, model.module.classifier2.parameters()))
+ list(map(id, model.module.classifier3.parameters()))
)
base_params = filter(lambda p: id(p) not in ignored_params, model.parameters())
optimizer_ft = optim.SGD([
{'params': base_params, 'lr': 0.1 * opt.lr},
{'params': model.module.classifier0.parameters(), 'lr': opt.lr},
{'params': model.module.classifier1.parameters(), 'lr': opt.lr},
{'params': model.module.classifier2.parameters(), 'lr': opt.lr},
{'params': model.module.classifier3.parameters(), 'lr': opt.lr},
], weight_decay=5e-4, momentum=0.9, nesterov=True)
else:
model = model.cuda()
if not opt.CPB:
ignored_params = list(map(id, model.classifier.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params, model.parameters())
optimizer_ft = optim.SGD([
{'params': base_params, 'lr': 0.1 * opt.lr},
{'params': model.classifier.parameters(), 'lr': opt.lr}
], weight_decay=5e-4, momentum=0.9, nesterov=True)
else:
ignored_params = (list(map(id, model.classifier0.parameters()))
+ list(map(id, model.classifier1.parameters()))
+ list(map(id, model.classifier2.parameters()))
+ list(map(id, model.classifier3.parameters()))
)
base_params = filter(lambda p: id(p) not in ignored_params, model.parameters())
optimizer_ft = optim.SGD([
{'params': base_params, 'lr': 0.1 * opt.lr},
{'params': model.classifier0.parameters(), 'lr': opt.lr},
{'params': model.classifier1.parameters(), 'lr': opt.lr},
{'params': model.classifier2.parameters(), 'lr': opt.lr},
{'params': model.classifier3.parameters(), 'lr': opt.lr},
], weight_decay=5e-4, momentum=0.9, nesterov=True)
if opt.adam:
optimizer_ft = optim.Adam(model.parameters(), opt.lr, weight_decay=5e-4)
exp_lr_scheduler = lr_scheduler.MultiStepLR(optimizer_ft, milestones=[60 - start_epoch, 75 - start_epoch], gamma=0.1)
######################################################################
dir_name = os.path.join('/home/ubuntu-guangzhaodai/Desktop/AICIty-reID-2020/data/outputs', name)
if not opt.resume:
if not os.path.isdir(dir_name):
os.mkdir(dir_name)
# record every run
copyfile('./train_2020.py', dir_name + '/train.py')
copyfile('./model.py', dir_name + '/model.py')
# save opts
with open('%s/opts.yaml' % dir_name, 'w') as fp:
yaml.dump(vars(opt), fp, default_flow_style=False)
if fp16:
# model = network_to_half(model)
# optimizer_ft = FP16_Optimizer(optimizer_ft, dynamic_loss_scale=True)
model, optimizer_ft = amp.initialize(model, optimizer_ft, opt_level="O1")
if opt.angle:
criterion = AngleLoss()
elif opt.arc:
criterion = ArcLoss()
else:
criterion = nn.CrossEntropyLoss()
print(model)
model = train_model(model, criterion, optimizer_ft, exp_lr_scheduler,
start_epoch=start_epoch, num_epochs=80)
@layumi Hi. May I ask you something again? I would like to use my own previously trained model as a pre-trained model.
Let's say I have SE_imbalance_s1_128_p0.5_lr2_mt_d0_b24+v+aug_20220701 already trained model. And I would like to use my trained model as a pre-trained model for my next training. How could I do this?
I have tried as following:
python train_2020.py --data_dir ../../data/reid_data --name SE_imbalance_s1_128_p0.5_lr2_mt_d0_b24+v+aug --warm_epoch 5 --droprate 0 --stride 1 --erasing_p 0.5 --autoaug --inputsize 128 --lr 0.02 --use_SE --gpu_ids 0,1 --train_virtual --batchsize 128 --init_name SE_imbalance_s1_128_p0.5_lr2_mt_d0_b24+v+aug_20220701
however I got following error:
ModuleAttributeError("'{}' object has no attribute '{}'".format( torch.nn.modules.module.ModuleAttributeError: 'SENet' object has no attribute 'conv1'
Thanks
楼主，你好，刚开始入门reid，看了你的代码，最后的评估evaluate_result中的score是算的两个特征矩阵乘积，score = torch.mm(query_feature, torch.transpose( gallery_feature, 0, 1))，不是很理解，一个query的feature与gallery的features做度量，为什么要用乘法呀？求指教
@郑博士 你好。我这边运行train+virtual数据集，当运行到40个Epoch的时候，太慢了，我把它停了，然后跑测试集，但结果rank@5才不到90%，mAP才0.35，这正常吗？？会是哪里出错了呢？？
请问一下AIC2020官方提供的AIC20_track2_reid.zip和AIC20_track2_reid_simulation.zip数据集是什么关系呀？
为什么不是两个合在一起的呢？
是两个一起用，还是只用一个就行了？
郑博士你好：
我使用arc loss训练，网络不收敛。crossentropy 则正常。请问该工程arc loss、angle loss是否完善，可用？
@layumi 郑博士你好，请问一个epoch得运行多长时间，我这里运行了一个多小时，都没运行完？？是真的需要运行这么久，还是出错了？？
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.