zhunzhong07 / ecn Goto Github PK

Hi,thanks for giving such a fabulous code.Here are some questions about the code.


Both tables have the same result.
1、These two tables respectively show the results of this paper compared with unsupervised and supervised models, right?
2、Is the code in this article unsupervised or supervised？

thank u

Hi, thanks for your open source code, I am a bit confused about camstyle.I trained the duke dataset with stargan4reid.I found the image size with 256x128(h x w).And I also downloaded your trained camstyle images on BaiduYun.The image size is 256x256(h x w).Results are as follows.

python main.py -s market -t duke --logs-dir logs/market2duke-ECN

duke(camstyle you provided): mAP: 39.4%, top1: 63.2%
duke(I trained camstyle with hhl): mAP:29.9%, top1: 50.2%

Then I know that you retrained a new camstyle model.Could you release the training parameters? Thanks.

Hi,
I was trying main.py and encountered the following error.

/usr/local/lib/python3.6/dist-packages/sklearn/externals/joblib/externals/cloudpickle/cloudpickle.py:47: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses
  import imp
log_dir= /home/zamra/PYreps/Person_ReId/ECN/logs
Namespace(arch='resnet50', batch_size=128, data_dir='/home/zamra/PYreps/Person_ReId/ECN/data', dist_metric='euclidean', dropout=0.5, epochs=60, epochs_decay=40, evaluate=False, features=4096, height=256, inv_alpha=0.01, inv_beta=0.05, knn=6, lmd=0.3, logs_dir='/home/zamra/PYreps/Person_ReId/ECN/logs', lr=0.1, momentum=0.9, output_feature='pool5', print_freq=1, re=0.5, resume='', source='duke', target='market', weight_decay=0.0005, width=128, workers=8)
DA dataset loaded
  subset   | # ids | # images
  ---------------------------
  source train    |     0 |        0
  target train    | 'Unknown' |        0
  query    |     0 |        0
  gallery  |     0 |        0
Traceback (most recent call last):
  File "main.py", line 229, in <module>
    main(args)
  File "main.py", line 94, in main
    args.re, args.workers)
  File "main.py", line 50, in get_data
    shuffle=True, pin_memory=True, drop_last=True)
  File "/home/zamra/.local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 802, in __init__
    sampler = RandomSampler(dataset)
  File "/home/zamra/.local/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 64, in __init__
    "value, but got num_samples={}".format(self.num_samples))
ValueError: num_samples should be a positive integeral value, but got num_samples=0

I want to use ECN for my own dataset, but it lecks the camstyle file, how to generate the camstyle file?

Thank you for your open source code，In your paper,you analysed the computational cost of the exemplar memory.But when I ran your code,I found It took me more than two hours in my machine. so I want to ask you whether you run on a single GPU or multiple GPUs ? Or do you have any way to reduce the running time? By the way,my machine configuration is a 12G single GPU.

I look forward to your reply，Thank you!

Hello there

Is there a way that I use the losses other than camera invariance loss and keep the camera invariance to 0 ?

Hi, thank you for sharing. There are some part of the code i don't understand their importance. one_hot is defined as torch.ones(inputs.size(0), knn). So why do you need to multiply the confidence score weights by a matrix which values are all ones?

I try to to visualize the results, and there not difference between one_hot * weights and simply using weights.
Thank you

I noticed the performance is slightly improved compared to the baseline (Source Only) in your paper when you just use the exemplar invariance and source domain labeled data? However, I try to implement it , and the rank1(24%) is lower than that of baseline (28.9). Do you have some training tricks?Can you provide some suggestions?

For Duke to Market-1501 transfer task, I got only

Mean AP: 6.5%
CMC Scores
top-1 20.5%
top-5 32.9%
top-10 40.0%
top-20 47.7%
How can I get the same results as paper? Did you run this version on your machine? Please advice me.

Hi, it is happy to follow your papers again.

About the code, I am confused about several places.

Question1:

tgt_feat = F.normalize(x)
tgt_feat = self.drop(tgt_feat)
if output_feature == 'tgt_feat':
     return tgt_feat

norm-drop, so the tgt_feat is not norm, Does it matter when you solve the probability of $p(i|x_{t,i})$? Because this is the first time I see get the feature after drop to calculate.

Question2:

        fixed_names = []
        for name, module in self.base._modules.items():
            if name == "layer3":
                # assert fixed_names == ["conv1", "bn1", "relu", "maxpool", "layer1", "layer2"]
                break
            fixed_names.append(name)
            for param in module.parameters():
                param.requires_grad = False

fixed them is the result from experiment or other sources?

Qestion3:

what is mean of ones_mat * weights? Is it because of historical something?

        ones_mat = torch.ones(targets.size(0), k).to(self.device)
        targets = torch.unsqueeze(targets, 1)
        targets_onehot = torch.zeros(inputs.size()).to(self.device)

        weights = F.softmax(ones_mat, dim=1)
        targets_onehot.scatter_(1, index_sorted[:, 0:k], ones_mat * weights)

Question4:

targets_onehot.scatter_(1, targets, float(1))

How do you guarantee that $x_{t,i}$ is not in $M(x_{t,i})$. I mean, my understanding is weight of $M(x_{t,i})$ should be sum to be 1. Or there is something I am wrong?

Question5:

    def set_model_train(self):
        self.model.train()

        # Fix first BN
        fixed_bns = []
        for idx, (name, module) in enumerate(self.model.module.named_modules()):
            if name.find("layer3") != -1:
                # assert len(fixed_bns) == 22
                break
            if name.find("bn") != -1:
                fixed_bns.append(name)
                module.eval()

My understanding is that since you have fixed conv1~layer2, the bn between them should be mode of eval. Is that right?

Question6:

The CamStyle is the same as the HHL?

Thank you, sir.

I am looking for a ReID demo. thanks for open source. I found you didn't release pretrained model. So if convenient, please upload these pretrained wegiths . thank you

I'm getting RunTime Error: Legacy Autograd function with non-static forward method is depracted.
Please use new-style autograd function with static forward method.

Hi, Thanks for your excellent work.
It seems that there something wrong with the neighbor number in invariance loss.
In Line 53 of file reid/loss/invariance.py, loss = self.smooth_loss(inputs, targets) use no parameter neighbor number self.knn.
I think it should be loss = self.smooth_loss(inputs, targets, self.knn)
Can you check it?

Hi, how much memory do you need to run this code?

Hi!
Using the msmt17 dataset you provided in the readme file, there are 60.910 images for testing (only one directory in the provided link). While in the original dataset, the authors provide 82161 images.

Can you confirm that you are using the original splitting strategy?
This is what I get using the original dataset
bbox 126,441
identities 4101 i
train images 32,621
train id 1,041
testing images 82,161
testing id 3,060
query images 11,659