MART v0.6.1

What's Changed

We decouple the vision component from Adversary, so that users who work on audio adversary or text adversary do not need to install vision components any more.

By moving many dependency to optional groups, we make it flexible for developers to integrate MART into existing projects.

We add support to ImageNet-normalized input and [0,1] image input by adding batch canonicalization transforms and reverse transforms in configs.

The detailed changes are listed below.

Adversary

• Prepare target models before running attacks by @mzweilin in #249
• Add batch_c15n for image of [0,1] or imagenet-normalized by @mzweilin in #248
• Add utils for config instantiation. by @mzweilin in #250
• Add mart.nn.Get() to extract a value from the kwargs dict. by @mzweilin in #251

Dependency

• Upgrade flake8 to 7.0.0 in pre-commit by @mzweilin in #245
• Upgrade pydantic to avoid CVE-2024-3772 by @mzweilin in #253
• Make dependencies flexible with optional groups by @mzweilin in #254
• Make mart importable with the core dependency group alone by @mzweilin in #255
• Update dependency by @mzweilin in #252

Development and CI

• Specify read-all permissions in workflows. by @mzweilin in #246
• Hash-pin dependency in GitHub Actions by @mzweilin in #247
• Skip object detection tests if pycocotools is not installed. by @mzweilin in #257

Full Changelog: v0.6.0...v0.6.1

v0.6.0

What's Changed

Dataset

• Add FiftyOne datamodule by @lumurillo in #209
• Enhance FiftyOne datamodule on bbox-only datasets by @mzweilin in #234

Adversary

• Tolerate tiny errors in enforcer. by @mzweilin in #212
• Make reusable attack config components by @mzweilin in #211
• Make adversarial training callback work by @mzweilin in #210
• Add Adversary.batch_c15n by @mzweilin in #215
• Move batch_c15n out of Adversary by @mzweilin in #220
• Add composable functions to Composer by @mzweilin in #228
• Fix adversarial image visualizer with canonical batches by @mzweilin in #227
• Move Perturber to Composer.Perturber by @mzweilin in #231
• Add learning rate schedulers in Adversary by @mzweilin in #233
• Make mask key configurable. by @mzweilin in #225
• Allow trainer.precision options in Adversary. by @mzweilin in #235
• Modular Composer by @mzweilin in #240
• Add unbounded patch adversary for object detection by @mzweilin in #241
• Add Lp-bounded patch adversary for object detection by @mzweilin in #242
• Simulation-in-loop adversary by @mzweilin in #243
• Visualize intermediate images of Composer by @mzweilin in #244

Misc

• Disable FiftyOne's tracker. by @mzweilin in #223
• Add a YAML config file generator by @mzweilin in #224
• Add python-fire in MART's dependency. by @mzweilin in #236
• Upgrade pre-commit flake8 to 6.0.0 by @mzweilin in #239

Full Changelog: v0.5.3...v0.6.0

MART v0.5.3

What's Changed

• Update dependency on PyTorch >=2.0.1 by @mzweilin in #237

Full Changelog: v0.5.2...v0.5.3

MART v0.5.2

What's Changed

• Upgrade dependency to torchmetrics == 1.0.1 by @mzweilin in #205

Full Changelog: v0.5.1...v0.5.2

MART v0.5.1

What's changed

• Revert "Make input a dictionary for multi-modal object detection" #95

Full Changelog: v0.5.0...v0.5.1

MART v0.5.0

What's Changed

Improve Adversary using LightningModule and Trainer from Lightning

• Configuring optimization hyper-params in Perturber by @mzweilin in #91
• Remove NoAdversary by @dxoigmn in #102
• Decorate initializers __call__ with torch.no_grad instead of using tensor.data by @dxoigmn in #107
• Decorate projectors __call__ with torch.no_grad instead of using tensor.data by @dxoigmn in #108
• Enforcing threat models by @mzweilin in #92
• Make IterativeGenerator a Trainer by @dxoigmn in #113
• Remove BatchComposer in favor of batch-aware Composer by @dxoigmn in #121
• Remove BatchEnforcer in favor of batch-aware Enforcer by @dxoigmn in #122
• Make Enforcer modality-aware by @mzweilin in #124
• Add MaskAdditive composer. by @mzweilin in #129
• Make Projector batch aware by @dxoigmn in #120
• Update GradientModifier to be an in-place operation by @dxoigmn in #123
• Replace tuple with Iterable[torch.Tensor] by @dxoigmn in #134
• Better Perturber by @dxoigmn in #146
• Better Optimizer by @dxoigmn in #147
• Adversary as pl.LightningModule by @dxoigmn in #103
• Improve Adversary's progress bar by @mzweilin in #161
• Move load_state_dict into LitModular by @dxoigmn in #145
• Allow LitModular to load from checkpoints with states of Adversary by @mzweilin in #160
• Add gradient monitor callback by @dxoigmn in #144
• Add Image initializer by @dxoigmn in #142
• Make the Adversary Trainer accept loggers. by @mzweilin in #164
• Hide parameters of Adversary's Perturber from DDP by @mzweilin in #166

Improve LitModular and mart.nn

• Allow modules to skip CallWith by @dxoigmn in #106
• Allow LitModular.*_step_log to be a dictionary by @dxoigmn in #168
• Let loss/preds/target output keys be configurable in LitModular by @dxoigmn in #169
• Remove *_step_end from LitModular by @dxoigmn in #170
• Make SequentialDict return outputs from all modules as DotDict by @dxoigmn in #171
• Don't lookup non-str keys in SequentialDict by @dxoigmn in #174
• Enable overriding of special args in SequentialDict.forward by @dxoigmn in #175
• Add _train_mode_ and _inference_mode_ special args to SequentialDict by @dxoigmn in #176
• Add _call_ special arg to SequentialDict by @dxoigmn in #184
• Allow LitModular's load_state_dict to accept a str by @dxoigmn in #189

Update dependency

• Upgrade to PyTorch 2.0 by @mzweilin in #192

Improve performance

• Speedup adversarial training with a simple perturber by @mzweilin in #89

Move non-essential functionalities to examples

• Fix the autoattack wrapper by @mzweilin in #80
• Remove test on RobustBench. by @mzweilin in #179
• Purge tests on RobustBench by @mzweilin in #181
• Move RobustBench and Carla stuff from mart to examples by @mzweilin in #183
• Move ART to extras by @dxoigmn in #186
• Move ART Wrapper to examples by @mzweilin in #195

Improve datamodule

• Make input a dictionary for multi-modal object detection by @mzweilin in #95
• Quantize datamodule input by @mzweilin in #112

Miscellaneous

• Fix source header license by @mzweilin in #99
• Fix source header license in tests by @mzweilin in #100
  Full Changelog: v0.4.1...v0.5.0

MART v0.4.1

What's Changed

• Move ART to extras by @dxoigmn in #186

Full Changelog: v0.4.0...v0.4.1

v0.4.0

What's Changed

New features

• Allow adversaries to run at any layer. by @mzweilin in #72

Misc.

• Remove Python 3.10 from test workflow. by @mzweilin in #67
• Upgrade to PyTorch 1.13.1. by @mzweilin in #75
• Upgrade isort to 5.12.0 by @dxoigmn in #82

Full Changelog: v0.3.1...v0.4.0

MART v0.3.1

What's Changed

Bug fixes

• Fix a Callable type annotation. by @mzweilin in #66

Full Changelog: v0.3.0...v0.3.1

You can install v0.3.1 by running $ pip install https://github.com/IntelLabs/MART/archive/refs/tags/v0.3.1.zip.

MART v0.3.0

What's Changed

New features

• Composable model sequences that are easy to change in command line. (#55)
• Add more options to profilers. (#20)

Bug fixes

• Multi-GPU training. (#24)
• Allow —resume=? without other parameters. (#46)
• Add GroupNorm32 to namespace mart.nn. (#45)
• Capture exceptions with empty error messages. (#61)

Misc.

• Type annotations. (#29)
• Return stdout for debugging failed tests. (#62)
• Automatically installing developer's dependency. (#59)
• Update installation instructions. (#21)

Full Changelog: v0.2.1...v0.3.0

You can install v0.3.0 by running $ pip install https://github.com/IntelLabs/MART/archive/refs/tags/v0.3.0.zip.