v0.2.3

Merge in rebuttal stuff on spectral analysis.

v0.2.2

Address minor issues in #25.

v0.2.1

• The codebase now supports the following additional models

  • BartForConditionalGeneration (when positional embedding layers are frozen)
  • T5ForConditionalGeneration
  • OPTForCausalLM
  • ViTForImageClassification (when isolated parameters are frozen; see this example)
  • DeiTForImageClassification (when isolated parameters are frozen)
  • BeitForImageClassification (when isolated parameters are frozen)

• Minor rewrite of the privacy engine, sharing code between ghost and default clipping, thus making future maintaining potentially easier.

• Add example file for fine-tuning of vision Transformers on CIFAR-10.

• Use callback style implementation for playing with params and gradients after clipping.

• Allow forward and backward hooks to use additional args and kwargs (not necessarily torch.Tensors).

• Rewrite privacy accounting and remove GDP + CLT, since it under accounts.

v0.1.0

Initial commit.