amuamushu / adv_avod_ssn

Deep Sensor Fusion for Single Source Robustness
https://ayushmore.github.io/2022-03-07-improving-robustness-via-adversarial-training/
0 stars 0 forks source link

Find code differences between trainsin and train. #32

Closed amuamushu closed 2 years ago

amuamushu commented 2 years ago

We plan on formatting our adversarial training to be similar to trainsin since it would make incorporating adversarial examples easier.

To make sure we are implementing our adversarial training correctly, we need to figure out what are the differences between the two.

amuamushu commented 2 years ago

Main difference in trainsin config:

image
amuamushu commented 2 years ago

A difference between how Taewan and we perturb the images is Taewan perturbs before running the image through the model. It doesn't seem like he saves the perturbed image itself, rather he saves the properties of the perturbed image in the feed dictionary.

We would have to perturb as the model is running because we perturb on the image rois.