Cory Cornelius
Cory Cornelius
# What does this PR do? This PR implements the "dual mode" of using torchvision's FasterRCNN module using configuration, rather than code. This PR depends upon: - [x] #171 -...
# What does this PR do? I think it's worth thinking about whether we should merge this PR. Note that `AttackInEvalMode` will be wrong in newer versions of PL since...
# What does this PR do? Writing images to disk a bad idea when we can write images to Tensorboard. This PR also genericizes PerturbedImageVisualizer into a configurable ImageVisualizer so...
# What does this PR do? A big missing feature of MART is the ability to generate universal perturbations. This PR tries to change that. The way universal perturbations should...
# What does this PR do? - [x] Make Composer an nn.Module. - [ ] Think more carefully about Composers vis-a-vis universal composers versus sample-specific composers. - [ ] Make...
Right now MART does two things: 1. There is a "outer" training loop that more-or-less executes the standard training loop that takes a model and fits/finds some parameters. 2. There...
Torchvision's v2 transforms support for transform images and bounding boxes similar to how we implement via ExTransform: https://github.com/IntelLabs/MART/blob/20d20789548c232e8d9a13079c2cef222fb9889b/mart/transforms/extended.py#L35 I think we should just standardize on Torchvision's implementation. https://pytorch.org/vision/main/auto_examples/plot_transforms_v2.html
Why do attack objectives and gains use `CallWith`? https://github.com/IntelLabs/MART/blob/9d1a03d41e8162bfa4590bf600d4d1a19854262e/mart/configs/attack/gain/modular.yaml#L1 https://github.com/IntelLabs/MART/blob/9d1a03d41e8162bfa4590bf600d4d1a19854262e/mart/configs/attack/objective/object_detection_missed.yaml#L1 Shouldn't we just be able to pass these arguments in using sequences? Or what is going on here?
Some options: - https://github.com/Tianxiaomo/pytorch-YOLOv4 - https://github.com/AlexeyAB/Yet-Another-YOLOv4-Pytorch (forked from https://github.com/VCasecnikovs/Yet-Another-YOLOv4-Pytorch) I'm partial to the AlexeyAB repo.
Model authors sometimes use `nn.Module.training` to change the control flow of their model. This is problematic because we often make the assumption a model in training more or less produces...