dp-custom

KO | EN

gitlite — search

#javascript #hacktoberfest #typescript #android #machine-learning #docker #nodejs #macos #java #awesome #kubernetes #awesome-list #vue #openai

(★ 9)

dp-custom

dp-custom Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# dp-custom
To quickly reproduce results for DP fine-tuning on CIFAR10, just install the following libraries via pip;
  - pytorch
  - numpy
  - torchvision
  - timm
  - fastdp from https://github.com/awslabs/fast-differential-privacy (preferred to opacus for a number of reasons)
  - opacus (not preferred)
  - tqdm
  - prv_accountant

And then just cd into dp_finetuning and run `python finetune_classifier_dp.py`. To run the accounting (which can be quite slow) pass `sigma=-1`. There are a number of command-line arguments as described below.

To run hyperparameter tuning with the linear scaling rule, run `python linear_scaling.py`.

For more general experiment reproduction, run the following command:

```
python {script}.py\
    --dataset_path ${0}\
    --dataset ${1}\
    --lr ${2}\
    --epsilon ${3}\
    --epochs ${4}\
    --arch ${5}\
```

For conventional CV experiments (ImageNet, CIFAR10, CIFAR100, FashionMNIST, EMNIST, MNIST, STL10, SVHN) `{script} = finetune_classifier_dp`. If there are any errors during extracting features please raise a comment to the authors -this

For OOD experiments in Wilds (waterbirds, fmow, domainnet, camelyon) `{script} = wilds_finetune_classifier_dp` after following directions in the cited papers to download and split the datasets accordingly.

For OOD experiments transferring from CIFAR10/CIFAR100 first add the `"--save_weights"` flag when running a conventional CV experiment, then use `{script} = test_transfer` and specify the transfer dataset with `--transfer_dataset ${6}`. Make sure to download CIFAR10C/CIFAR100C from the cited paper.

Dependencies (non-exhaustive):
  - pytorch
  - numpy
  - torchvision
  - opacus
  - timm
  - tqdm
  - fastdp

dp-custom

dp-custom Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

dp-custom

dp-custom Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README