cdluminate / robrank Goto Github PK

a metric learning reality check, eccv 2020

in defense/*.py

margin=configs.triplet.margin_euclidean if model.lossfunc._metric in ('E', 'N')

Hi, Thanks for your great work!

When I run the command 'python3 bin/train.py -C mnist:rc2f2:ptripletN --do_test', I got the output log as follows:

I don't make any change of the original code. Could you give me some advice about it?

My mainly environment:
Python: 3.8.5 (anaconda)
PyTorch: 1.8.1+cu111
GPU: Nvidia RTX3090 (24GB)

Thanks

I should document it

After running "python3 bin/train.py -C mnist:rc2f2:ptripletN", I find model.config.maxepoch is 8. It's different from the experiment in the paper. It was written that the C2F2 model is trained with the Adam optimizer for 16 epochs. Then I notice the code "self.maxepoch //= 2" of rc2f2 function in configs_rank.py. I try to change the code as follows. It doesn't work, and model.config.maxepoch = 8 . How can I change the epoch into 16?

Hope you can help me.
Thanks and Regards.

Hi, I'm trying to do a defensive training on resnet18 without SPC2 term so that I can futher alter the loss function.
However, I'm constantly encountering with the following problem, but I have no clue about where could possbily be wrong.

Thanks!

currently it takes 2x video memory for act training compared to standard training process. Can we reduce the memory requirement so that the code can be run on more budget GPUs?

Hi Mo,

I am reading your new paper "Enhancing Adversarial Robustness for Deep Metric Learning". It is quite interesting and the performance is good! May I ask is there any code for it? Or do you prepare to update the code for the new method HM? :)
Thanks.

Best regards,
Cosmo

Accordingly, we should move the module not found error in init.py into an exception when getting attr from models.

Excuse me, what's the value of 'm‘ in “Batched Adversarial Attack” of your paper? Could you give a detailed guidance to reproduce the result of Table in your paper?

Hi Mo,

What command do you use to obtain the metrics (R@1, mAP, NMI) on benign examples?

I've used your code to train a robust retrieval model (cub-rres18ghmetsm-ptripletN) for the CUB dataset, and at the end I get validation accuracy similar to the one you reported in your CVPR '22 paper (taken from logs):

Validation │ r@M= 18.3 r@1= 0.376 r@2= 0.485 mAP= 0.172 mAP@R = 0.222 NMI= 0.485

The problem arises, when I try to reload the checkpoint through either validate or continue training, I get very different results:

Validation │ r@M= 6.8 r@1= 0.450 r@2= 0.584 mAP= 0.164 mAP@R = 0.239 NMI= 0.546

The problem is that I get very similar "weird" results no matter what checkpoint I pass into the Train / Validate Class in cmdline.py.

Thanks!

https://robustbench.github.io/

prevalent in the classification defense field.