Semantic Robustness #259
Semantic Robustness #259
Conversation
…local configs folder.
mart/callbacks/anomalib.py
Outdated
| best_batch = None | ||
| batch_history = defaultdict(list) | ||
|
|
||
| for step in (pbar := trange(self.steps, desc="Attack", position=1)): |
There was a problem hiding this comment.
This := is a bit of an exotic syntax, doesn't seem to appear anywhere else in the MART repo
There was a problem hiding this comment.
There's this:
MART/mart/callbacks/progress_bar.py
Line 16 in 36609c7
Not sure how helpful it is to re-use
There was a problem hiding this comment.
Ah you should learn about the walrus (:= a Python 3.8 feature)! It's super useful in some circumstances.
I think using ProgressBar will be better once this code becomes MART native.
There was a problem hiding this comment.
Good article here about the walrus operator: https://realpython.com/python-walrus-operator/
mart/callbacks/anomalib.py
Outdated
| image_hsv = rgb_to_hsv(image_adv) | ||
| image_hue, image_sat, image_val = torch.unbind(image_hsv, dim=-3) | ||
|
|
||
| # Additive hue perturbation with STE clipping |
There was a problem hiding this comment.
Comment is inaccurate, we never did STE for torch.remainder. I think it's fine though to not do it, should just be a comment reword
mart/callbacks/anomalib.py
Outdated
| image_sat = image_sat + sat[:, None, None] | ||
| image_sat = image_sat + (torch.clip(image_sat, 0.0, 1.0) - image_sat).detach() | ||
|
|
||
| # Convert image fro HSV to RGB space |
mart/callbacks/anomalib.py
Outdated
|
|
||
| loss = negative_loss + positive_loss | ||
|
|
||
| # decrease negatives and increase positives |
There was a problem hiding this comment.
Given that you removed maximize=True in Adam, it may be worth adding an extra comment here that this will be minimized, and the intended effect is exactly opposite than that of the comment.
There was a problem hiding this comment.
I assume losses are made to be minimized and gains are made to be maximized, but fixed in 85e85bb, I think.
|
@mzweilin: Would be nice to get this merged. Really need help on how to package this. |
|
LGTM! Thanks for the README updates! |
What does this PR do?
This PR implements a semantic robustness callback for anomalib as described in https://arxiv.org/abs/2405.07969. We make one critical design choice different from the paper: we supply the model with an augmented image and mask for computing metrics. Additionally, anomalib does not implement patching, so all images are square by default. We believe these changes shouldn't significantly alter results since 1) rotating masks is fine so long as we impute pixel mask values as "non-anomalous" or 0 valued.
Type of change
Please check all relevant options.
Testing
Please describe the tests that you ran to verify your changes. Consider listing any relevant details of your test configuration.
pytestCUDA_VISIBLE_DEVICES=0 python -m mart experiment=CIFAR10_CNN_Adv trainer=gpu trainer.precision=16reports 70% (21 sec/epoch).CUDA_VISIBLE_DEVICES=0,1 python -m mart experiment=CIFAR10_CNN_Adv trainer=ddp trainer.precision=16 trainer.devices=2 model.optimizer.lr=0.2 trainer.max_steps=2925 datamodule.ims_per_batch=256 datamodule.world_size=2reports 70% (14 sec/epoch).Before submitting
pre-commit run -acommand without errorsDid you have fun?
Make sure you had fun coding 🙃