private-transformers
This codebase facilitates fast experimentation of differentially private training of Hugging Face transformers.
What is this? Why an extra codebase?
- This codebase provides a privacy engine that builds off Opacus, but works way more smoothly with Hugging Face's transformers library.
- Additionally, we support the ghost clipping technique (see Section 4 of this preprint on how it works) which allows privately training large transformers with considerably reduced memory cost -- in many cases, almost as light as non-private training -- at a modest run-time overhead.
- With this codebase, we have fine-tuned very large pretrained models, yielding some of the best performing differentially private NLP models to date. Some of these models have performance matching strong non-private baseline approaches. We see strong empirical evidence that highly performant DP NLP models could be built on modest datasets.
Installation
Make sure you have python>=3.8; run the following command:
pip install git+https://github.com/lxuechen/private-transformers.git
To check the package is installed properly, be sure to run the test suite (requires pytest and a GPU) via the following command:
pytest -s tests
Usage
Basic usage
Privately training Hugging Face transformers with our codebase simply consists of 4 steps:
- Create your favourite transformer model and optimizer; attach this optimizer to a
PrivacyEngine - Compute a per-example loss (1-D tensor) for a mini-batch of data
- Pass the loss to
optimizer.steporoptimizer.virtual_stepas a keyword argument - Repeat from step 2
Below is a quick example:
import transformers, torch
from private_transformers import PrivacyEngine
import torch.nn.functional as F
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = transformers.GPT2LMHeadModel.from_pretrained('distilgpt2').to(device)
optimizer = torch.optim.Adam(params=model.parameters(), lr=1e-4)
privacy_engine = PrivacyEngine(
model,
batch_size=10,
sample_size=50000,
epochs=3,
max_grad_norm=0.1,
target_epsilon=3,
)
privacy_engine.attach(optimizer)
batch_size, seq_len = 10, 20
# Inputs are batch-first format, i.e., the first dimension of tensors must be batch dimension.
input_ids = torch.randint(size=[batch_size, seq_len], low=0, high=100, device=device)
# Calling `.train()` is very important; otherwise underlying forward and backward hooks don't run.
model.train()
outputs = model(input_ids=input_ids, return_dict=True)
labels = input_ids[:, 1:, ]
logits = outputs.logits[:, :-1, :].permute(0, 2, 1)
# `loss` is a 1-D tensor of shape (batch_size,).
loss = F.cross_entropy(logits, labels, reduction="none").mean(dim=1)
# This step is different from existing workflows:
# Don't call `loss.backward`; leave it to `optimizer.step` to handle backward.
optimizer.step(loss=loss)
The biggest differences compared to Opacus are:
- We require the per-example loss (a 1-D tensor) be passed into
optimizer.step(oroptimizer.virtual_step) - The per-example loss must be passed in as a keyword argument.
loss.backward()shouldn't be called on the user end; it's called internally inoptimizer.step( oroptimizer.virtual_step).- Inputs should be in batch-first format; there isn't a toggle to switch between different formats in the engine.
Ghost clipping: memory saving differentially private learning
Turning on ghost clipping requires changing only 1 line. You should notice a drastic reduction in peak GPU memory usage once this is turned on, at a potential cost of slower training speed. One might find this especially useful when constrained to only use older GPUs with small VRAMs or fitting super large models.
import transformers, torch
from private_transformers import PrivacyEngine
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = transformers.GPT2LMHeadModel.from_pretrained('distilgpt2').to(device)
optimizer = torch.optim.Adam(params=model.parameters(), lr=1e-4)
privacy_engine = PrivacyEngine(
model,
batch_size=10,
sample_size=50000,
epochs=3,
max_grad_norm=0.1,
target_epsilon=3,
ghost_clipping=True, # The only change you need to make!
)
privacy_engine.attach(optimizer)
We ran stringent numerical tests to ensure the double-backward implementation is correct. Check out files in the tests folder for more on this.
Examples
Code in the examples folder roughly reproduces our results for the table-to-text and classification tasks. There may be some minor discrepancies, since hyperparameters there aren't exactly what's used in the paper. Nevertheless, it should be sufficient to get things started. Detailed instructions are in the readme file of each subfolder.
Currently supported Hugging Face models
- OpenAIGPTLMHeadModel
- OpenAIGPTDoubleHeadsModel
- GPT2LMHead
- GPT2DoubleLMHead
- BertForSequenceClassification
- RobertaForSequenceClassification
- AlbertForSequenceClassification
Not all models in the Hugging Face library are supported. The main additional work here is to
- support per-example gradients for bespoke modules (e.g., T5LayerNorm), and
- ensure
position_idsare repeated.
We plan to support more models in the future if there's such a need. Feel free to open an issue if you may want to try out specific models that aren't in the current list.
FAQ
I wrote some answers to potential questions here.
Acknowledgements
It would have been impossible to develop this codebase without cool past works and existing codebases. We roughly follow the PrivacyEngine design in Opacus==0.13.0. We directly use an off-the-shelf package for tightly tracking tradeoff functions while composing multiple private mechanisms.
Disclaimer
- This codebase is not yet production-grade, e.g., cryptographically secure PRNGs are required for sampling noise -- our codebase currently does not use these strong PRNGs.
- This codebase is born out of the need to experiment with various things for differentially private NLP in rapidly succession. I've tried my best to write clean code, though parts of this codebase may be less tidy than I had hoped given the extremely tight timeline.
Citation
If you found this codebase useful in your research, please consider citing:
@misc{li2021large,
title={Large Language Models Can Be Strong Differentially Private Learners},
author={Xuechen Li and Florian Tramèr and Percy Liang and Tatsunori Hashimoto},
year={2021},
eprint={2110.05679},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
290 Dec 20, 2022
22 Nov 13, 2022
581 Dec 21, 2022
9 Dec 28, 2022
82 Dec 19, 2022
20 Nov 18, 2022
23 Sep 27, 2022
42 Dec 20, 2022
914 Jan 7, 2023
141 Dec 30, 2022
58 Dec 7, 2022
77.3k Jan 3, 2023
6.4k Jan 9, 2023
1.2k Jan 8, 2023
342 Nov 21, 2022