mangye16 / reid-survey Goto Github PK

Hello, thanks for your hard work. I have a quick question while inspecting the backbone. Why is this line of Relu activation in Resnet deleted?

self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
self.bn1 = nn.BatchNorm2d(64)
# self.relu = nn.ReLU(inplace=True)   # add missed relu
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

大神您好，
由于我刚开始接触ReID, 在看您的博客时有些地方不太明白，您讲到“通过可视化 rank list 可以快速定位模型的问题”。想问一下，有具体的可视化代码和图示效果吗？谢谢。

您好，在Non_local模块中您写的是self.inter_channels = reduc_ratio//reduc_ratio，请问这是特意设置成这样的嘛，还是说应该是self.inter_channels = in_channels//reduc_ratio ？

谢谢

Hi~
It is a nice job. The proposed WRT loss seems to be the improved version of the original Triplet loss. Did you carry out any ablation studies on the proposed AGW baseline for cross-modality visible-infrared Re-ID? If so, how much performance improved by the WRT comparing to the triplet?

Thanks

Traceback (most recent call last):
File "tools/main.py", line 111, in
main()
File "tools/main.py", line 66, in main
model.module.load_param(cfg.TEST.WEIGHT)
File "./modeling/baseline.py", line 166, in load_param
self.state_dict()[i].copy_(param_dict[i])
KeyError: 'module.base.conv1.weight'
(ReID-Survey-master) server@server-System-Product-Name:/media/server/A/mqq/ReID-Survey-master$

可以提供ImageNet的预训练权重吗？

In config->defaults.py, there is a package named yacs imported, but there is not in the code file now. I can't find it. How can I download it?

跑market1501数据集，全部默认配置，bz=64，单卡map有88.4，4卡的精度只有85.5，请教下各位大佬，这到底是怎么回事，或者应该怎么调整

why insset the non-local attention in this way？

您好，在WRT给负样本加权时，对AN取负但并没有去除其中的0，导致在max时取的不是最近的负样本，与正样本的加权策略不一致，是否就是这样设计的呢？

Hi, please helpme to resolve this error.

Traceback (most recent call last):
File "tools/main.py", line 110, in
main()
File "tools/main.py", line 57, in main
model = build_model(cfg, num_classes)
File ".\modeling_init_.py", line 7, in build_model
cfg.MODEL.GENERALIZED_MEAN_POOL, cfg.MODEL.PRETRAIN_CHOICE)
File ".\modeling\baseline.py", line 126, in init
self.base.load_param(model_path)
File ".\modeling\backbones\resnet_nl.py", line 176, in load_param
param_dict = torch.load(model_path)
File "C:\Users\MAHE\anaconda3\envs\myenv\lib\site-packages\torch\serialization.py", line 365, in load
f = open(f, 'rb')
FileNotFoundError: [Errno 2] No such file or directory: '../resnet50-19c8e357.pth'

mangye 您好：
我手边有一段视频
想搭配您的ReID
做行人重识别
该如何操作？

Hello, 按照步骤成功训练，训练完的model是放在log里了吗？如下图片：

然后按照Test的方法（修改模型位置和名字在log路径里的任何一个），会有报错，如下：

请问，这个问题出在哪里？谢谢！

Hello.
I want to know the similarity of two images.
Your code is the evaluation output.
I want to know the code to predict.
E.g. Given two image inputs
Can you tell if this is the same or not?
Or get each features?

FileNotFoundError: [Errno 2] No such file or directory: '/home/cgv841/.cache/torch/checkpoints/resnet50-19c8e357.pth'
为什么我在哪个文件都没有找到类似 cgv841的字样？这是哪里来的呢？

Thanks for sharing the repo! Will you also share the training code as well?

Hi, firstly i would like to say, thanks for this amazing work.

Im trying to use the agw baseline model as a feature extractor to ReID for deepsort. After adapting the code for my use case, i see that the reid performance during tracking was not as good as a smaller model. Im suspicious that somewhere i messed up the preprocessing / postprocessing.

So if im using the provided pretrained model (to be specific, the market one), is this pipeline correct:

1. Using BBOX info, crop a patch from an RGB image
2. resize to input shape 128 x 256
3. Normalize to [0,1] by dividing by 255
4. substract the mean[0.485, 0.456, 0.406] and then divide by std [0.229, 0.224, 0.225]
5. change to NCHW ordering, transform to torch tensor (not important to mention i think)
6. feed to model, get output
7. normalize the output by doing output / np.linalg.norm(output,axis=1)
8. Use euclidian distance for feature distance

作者，你好！

我发现 non_local.py 中的 self.inter_channels 的计算方式如下：

但是原论文中应该是：self.inter_channels = in_channels // reduc_ratio

请问这是特意这样设置的吗，如果是的话，是基于什么样的理由呢？

您好，我在训练partial_reid数据集时，精度一直达不到论文的要求，相差还是蛮大的，

The test feature is normalized
2021-03-05 16:19:15,585 reid_baseline INFO: Validation Results - Epoch: 110
2021-03-05 16:19:15,585 reid_baseline INFO: mINP: 34.9%
2021-03-05 16:19:15,585 reid_baseline INFO: mAP: 50.2%
2021-03-05 16:19:15,585 reid_baseline INFO: CMC curve, Rank-1 :57.7%
2021-03-05 16:19:15,585 reid_baseline INFO: CMC curve, Rank-3 :66.0%
2021-03-05 16:19:15,585 reid_baseline INFO: CMC curve, Rank-5 :72.7%
2021-03-05 16:19:15,585 reid_baseline INFO: CMC curve, Rank-10 :80.3%
The test feature is normalized
2021-03-05 16:19:16,431 reid_baseline INFO: Validation Results - Epoch: 110
2021-03-05 16:19:16,431 reid_baseline INFO: mINP: 54.8%
2021-03-05 16:19:16,431 reid_baseline INFO: mAP: 54.8%
2021-03-05 16:19:16,431 reid_baseline INFO: CMC curve, Rank-1 :42.9%
2021-03-05 16:19:16,431 reid_baseline INFO: CMC curve, Rank-3 :60.5%
2021-03-05 16:19:16,431 reid_baseline INFO: CMC curve, Rank-5 :65.5%
2021-03-05 16:19:16,431 reid_baseline INFO: CMC curve, Rank-10 :73.9%
配置：
MODEL:
PRETRAIN_CHOICE: 'imagenet'
PRETRAIN_PATH: '../resnet50-19c8e357.pth'
CENTER_LOSS: 'on'
CENTER_FEAT_DIM: 2048
NAME: 'resnet50_nl'
WEIGHT_REGULARIZED_TRIPLET: 'on'
GENERALIZED_MEAN_POOL: 'on'

INPUT:
IMG_SIZE: [256, 128]
PROB: 0.5 # random horizontal flip
RE_PROB: 0.5 # random erasing
PADDING: 10

DATASETS:
NAMES: ('market1501')

DATALOADER:
PK_SAMPLER: 'on'
NUM_INSTANCE: 4
NUM_WORKERS: 8

SOLVER:
OPTIMIZER_NAME: 'Adam'
MAX_EPOCHS: 120
BASE_LR: 0.00035

CENTER_LR: 0.5
CENTER_LOSS_WEIGHT: 0.0005

WEIGHT_DECAY: 0.0005
IMS_PER_BATCH: 64

STEPS: [40, 70]
GAMMA: 0.1

WARMUP_FACTOR: 0.01
WARMUP_ITERS: 10
WARMUP_METHOD: 'linear'

CHECKPOINT_PERIOD: 40
LOG_PERIOD: 20
EVAL_PERIOD: 10

TEST:
IMS_PER_BATCH: 128
RE_RANKING: 'off'
WEIGHT: "path"
FEAT_NORM: 'on'
EVALUATE_ONLY: 'off'
PARTIAL_REID: 'on'

OUTPUT_DIR: "../log/market1501/Experiment-AGW-baseline"

_C.DATASETS.ROOT_DIR = ('../dataset')

dataset文件夹下面有market1501数据集、partial_reid数据集、partial_ilids数据集

数据集是在给出的链接上下载得到的

请教大佬，这个是由于什么原因造成的呀？非常感谢！