cea-list / adv-reid Goto Github PK

excuse me, these defense results are all under classify model?

After training the defended model GOAT, when i am attacking this model, it is giving accuracy as
rank1 : mean = 0.0008907363517209888, max = 0.0008907363517209888, min = 0.0008907363517209888, std = 0.0
rank5 : mean = 0.0029691210947930813, max = 0.0029691210947930813, min = 0.0029691210947930813, std = 0.0
rank10 : mean = 0.00801662728190422, max = 0.00801662728190422, min = 0.00801662728190422, std = 0.0
map : mean = 0.0020973790522167803, max = 0.0020973790522167803, min = 0.0020973790522167803, std = 0.0
Total computing time : 161.57796907424927 sec

Whereas, as stated in the paper, the accuracy should be higher after attacking the defended model.

Please provide me the checkpoints of defended model.

Thanks and Regards
Astha Verma
PhD Scholar

I have a confusion about the Equation, is the minus sign here a clericcal error? According your illustration,it seems that the sign should be a plus.
All the best.

could you tell me how to solve this problem?
The following is the order I executed:
python train.py
‘/data/GYP/code/CVPRW20_adv-reid/output’
‘../../data/market1501/’
--dataset 'market'
--gpu 1
--lr 0.0003
-wd 0
-n 100
-f checkpoint_default.py
-b 72
-ni 4
-e 2048
--pretrained
--triplet
--soft
--id_batch
I print the value:

shows that:

so I think the the program did not read the dataset properly

I have a problem to training a Classification model using the offline adversarial training defense method. I don't know what command to use.
Hope you can help me.
Thanks and Regards.

hello sir, if i choose the classification mdoel and use the SMA attack method, i understand that you use a backbone of resnet50 to extract features and a classifier to correspond the correct targets. And the model will return the feature of a image in the test mode, right?

However i notice that in the function "perturb _queries_self" which means the realization of SMA, you get the raw_features first, and then "image_adv = attack.perturb(data['image'].to(device), raw_features.to(device))". I see that in the advertorch code, the second argument of the function perturb should be the corresponding targets ,why do you use raw_features as input? i couldn't understand what does it mean. Wish for your reply ,thanks.

Hello, I got the following error while training Triplet model on Market-1501 dataset using the command provided in Training section of README.md:

I believe the error is caused by the following:

• https://github.com/CEA-LIST/adv-reid/blob/main/models/architectures/models.py#L42
• https://github.com/CEA-LIST/adv-reid/blob/main/models/architectures/models.py#L40

I think self.backbone.fc.in_features in these lines should be replaced with self.backbone.fc.out_features