Skip to content

Bringing latest changes into quantization fix#29

Merged
NightlyTwo58 merged 5 commits into
fix_quantizationfrom
fix_quant_merging
Jul 14, 2026
Merged

Bringing latest changes into quantization fix#29
NightlyTwo58 merged 5 commits into
fix_quantizationfrom
fix_quant_merging

Conversation

@NightlyTwo58

Copy link
Copy Markdown
Owner
  • most notably, CIFAR-100 and three models now
  • graphing utils improved
  • reorgs
  • TorchAO and STE clearly defined
  • caching performance improvements

NightlyTwo58 and others added 5 commits July 13, 2026 11:52
- centralizing graphing utils for the nth time
- make sure INT4 is visualized
- changed dataset to use CIFAR-100
- changed models to use three from TorchCV: ResNet56, WRN28_10, DenseNet100
- generalized more variables into config.py
- changed visualizations to remove any old PGD-STE plotting, changed PGD ablation graph to show BPDA and FP32 stats
- hopefully this is the last entry in PGD issues, merged the ablation and accuracy PGDs into the same utility
- old ablation PGD took worst of 5 runs
- new merged PGD uses static seeds
- formatting
- TorchAO was incorrectly integrated earlier such that (A) layers were detached and (B) quantization differed between hard and BPDA (STE-smoothed)
- quantization has now been factored out into its own file
- upgrades to graphing utilities; moved into own package
- committing "formaldata.py", used for condensed result plotting for paper
- implemeted unified framework for caching, increased reuse of results
- enabled a lot more attacks
@NightlyTwo58 NightlyTwo58 merged commit 35137c8 into fix_quantization Jul 14, 2026
2 checks passed
@NightlyTwo58 NightlyTwo58 deleted the fix_quant_merging branch July 14, 2026 06:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant