chenhao2345 / ice Goto Github PK

Hello, it's a great honor to read your excellent work, but when reproducing your paper code, I have the following questions: 1. Which part of the paper does "loss_ccl" in trainers.py correspond to? This loss function does not seem to improve the effect of the model. 2. At present, the highest accuracy of my reproduction is only 76.7%. How can I achieve the accuracy pointed out in the paper?

Thank you for your answer!

Is the number of GPUs or type of GPU effective in accuracy that achieved in the paper?
If I don't have 4 GPUs and I have 1 GPU, what would happen when running the code?
Is it possible that we cannot get the results of the paper?

Why does the performance degrade when I use only 1 GPU with all the same hyper-parameter?

How should I change the parameters when using 1 GPU.

Hello. Thank you for your great work.

How can I set the camera-aware or camera-agnostic during training?

For example, I do not want to use the camera ID in your method.

Is camera-agnostic not using camera ID at all??

Dear author, thank you for your outstanding work!

I have an immature doubt about the code implementation process:

I found that the calculation of ccloss: "f_out_t1" and "centers" do not take "l2 norm", that is, they do not represent cosine similarity. I want to ask, is there any special reason for the handling here?

Thank you for your coming reply!

Hi, I am getting error when I tried to train your network with
ResNet50_IBN.

I gave the arguments like this
"python examples/unsupervised_train.py --dataset-target market1501 --a
rch resnet_ibn50a"

File "examples/unsupervised_train.py", line 94, in create_model
model_1 = models.create(args.arch, num_features=args.features, dropout=args.dropout, num_classes=0)
File "/root/workplace/re-Id/ICE/ice/models/init.py", line 54, in create
return __factory[name](*args, **kwargs)
File "/root/workplace/re-Id/ICE/ice/models/resnet_ibn.py", line 130, in resnet_ibn50a
return ResNetIBN('50a', **kwargs)
File "/root/workplace/re-Id/ICE/ice/models/resnet_ibn.py", line 29, in init
resnet = ResNetIBN.__factorydepth
File "/root/workplace/re-Id/ICE/ice/models/resnet_ibn_a.py", line 187, in resnet50_ibn_a
state_dict = torch.load(model_urls['ibn_resnet50a'], map_location=torch.device('cpu'))['state_dict']
File "/root/anaconda3/envs/ICE/lib/python3.8/site-packages/torch-1.7.0-py3.8-linux-x86_64.egg/torch/serialization.py", line 581, in load
with _open_file_like(f, 'rb') as opened_file:
File "/root/anaconda3/envs/ICE/lib/python3.8/site-packages/torch-1.7.0-py3.8-linux-x86_64.egg/torch/serialization.py", line 230, in _open_file_like
return _open_file(name_or_buffer, mode)
File "/root/anaconda3/envs/ICE/lib/python3.8/site-packages/torch-1.7.0-py3.8-linux-x86_64.egg/torch/serialization.py", line 211, in init
super(_open_file, self).init(open(name, mode))
FileNotFoundError: [Errno 2] No such file or directory: '/home/hchen/Projects/Baseline/logs/pretrained/resnet50_ibn_a.pth.tar'

How can I load the pretrained weights of ResNet50-IBN?

Hi,

When I run the train on Market1501, the performance degrades.

Should DBSCAN clustering hyperparameter eps=0.5 in market1501?
, eps=0.55 for duke and eps=0.6 for msmt dataset?

您好，请问您目前的代码是达到82.3的那个版本吗，我试了好几次都差了七八个点。

Hi, I have some question about the implementation.

Why do you shuffle and reverse the mini-batch order before feeding it into the
online encoder and momentum encoder like the above code??

why didn't the author update the camera-aware memory bank in an epoch?

Can you provide the download address of the pre-trained model?

There is no code to test.

sorry to bother you. I'd like to ask a question. when I train the model with loss_ccl only, I can't achieve a good result, it will break down soon.

Sorry to bother you again. I'd like to ask a question. What is the meaning of intra-camera loss and cross-camera loss?

Hello!

In your solution matrix of Jaccard distances is precomputed and takes up a lot of space in RAM. Are there ways to solve this problem? This question is relevant for large datasets.

Thanks!

The accuracy of the model I obtained is low, only 50.6% in MAP, which is far different from the data in the paper. Is the parameter setting in the code not optimal?

请问您可以上传t-SNE可视化相关的代码吗？应该怎么选取可视化的行人ID？

Hi, I am a graduate student who is conducting an experiment because I was impressed by your ICE model. I meet an error “raise StopIteration StopIteration”, I find that train_loader_target is none. Do you know why?

Hi, I am a graduate student who is conducting an experiment because I was impressed by your ICE model. If I want to check the performance of ICE model's Supervised model, should I use cross-camera loss (Camloss)?

Sorry to bother you ,after adding intra-camera loss,I find the accuracy diseases whichever combination I use .If possible ,would you please tell me the reason ?Thank you very much.