An Open-Source Framework for Paramter-Efficient Tuning (Delta Tuning).

---

Overview • Installation • Basic Usage • Docs • Performance •

Overview

OpenDelta is a toolkit for parameter-efficient tuning methods (we dub it as delta tuning), by which users could flexibly assign (or add) a small amount parameters to update while keeping the most paramters frozen. By using OpenDelta, users could easily implement prefix-tuning, adapters, Lora, or any other types of delta tuning with preferred PTMs.

• Our repo is tested on Python 3.=-0 and PyTorch 1.9.0. Lower version may also be supported.

• A demo of using Opendelta to modify the PLM (E.g., BART).

News

• 2022.10.14 Release v0.3.0. We make the usage of default configurations of each delta tuning methods (i.e., the position they are attached) more friendly! If a custom model has our supported models as submodules inside, the default configuration is also available. Other key changes can be seen in Update Log
• 2022.10.10 Merge a long-developed branch v0.2.4 into the master branch. Key updates are (1) the an example unifying the delta tuning paradigm and the prompt-tuning paradigm; (2) and support for Delta Center, whose webpage is still under construction. Details can be seen in Update Log
• 2022.03.24 We notice several bugs in Soft Prompt Tuning and Prefix Tuning, mainly due to their need to customize attention ids, token_type_ids, we are fixing it! Currently, please use the other methods since they are stabler and better in performance.
• 2022.03.20 Add a colab example to illustrate efficient training and space-saving multitask-serving.
• 2022.03.20 A new pip version released.
• 2022.02.16 Support regular expression in named-based addressing.

Installation

create a virtualenv (optional)

conda create -n opendelta_env python=3.8
conda activate opendelta_env

Using Pip

Install OpenDelta using pip as follows:

pip install opendelta

To play with the latest features, you can also install OpenDelta from the source.

Build from Source

git clone https://github.com/thunlp/OpenDelta.git
cd OpenDelta

Option 1: If you won't modify the code, run

python setup.py install

Option 2: If you want to modify the code or keep the repo updated by git clone, run

python setup.py develop

Tips

• If you want to use mirror for installing the packages, please change the index_url in setup.cfg

• If you encounter network error using setup.py, please firstly install the dependencies via

pip install -r requirements.txt && python setup.py develop

Must Try

The follow codes and comments walk you through the key functionality of OpenDelta.

# use tranformers as usual.
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
t5 = AutoModelForSeq2SeqLM.from_pretrained("t5-large")
t5_tokenizer = AutoTokenizer.from_pretrained("t5-large")
inputs_ids = t5_tokenizer.encode("Is Harry Poter wrtten by JKrowling", return_tensors="pt")
t5_tokenizer.decode(t5.generate(inputs_ids)[0]) 
# >>> '<pad><extra_id_0>? Is it Harry Potter?</s>'


# use existing delta models
from opendelta import AutoDeltaModel, AutoDeltaConfig
# use existing delta models from DeltaCenter
delta = AutoDeltaModel.from_finetuned("thunlp/Spelling_Correction_T5_LRAdapter_demo", backbone_model=t5)
# freeze the backbone model except the delta models.
delta.freeze_module()
# see the change
delta.log()


t5_tokenizer.decode(t5.generate(inputs_ids)[0]) 
# >>> <pad> Is Harry Potter written by JK Rowling?</s>


# Now save the delta models, not the whole backbone model to tmp
delta.save_finetuned(".tmp")
# >>>  The state dict size is 1.443 MB
# >>>  We encourage users to push their final and public models to delta center to share them with the community!


# reload the model from local url and add it to pre-trained T5.
t5 = AutoModelForSeq2SeqLM.from_pretrained("t5-large")
delta1 = AutoDeltaModel.from_finetuned(".tmp", backbone_model=t5)
import shutil; shutil.rmtree(".tmp") # don't forget to remove the tmp files. 
t5_tokenizer.decode(t5.generate(inputs_ids)[0]) 
# >>> <pad> Is Harry Potter written by JK Rowling?</s>

# detach the delta models, the model returns to the unmodified status.
delta1.detach()
t5_tokenizer.decode(t5.generate(inputs_ids)[0])  
# >>> '<pad><extra_id_0>? Is it Harry Potter?</s>'

# use default configuration for cunstomized wrapped models which have PLMs inside.
import torch.nn as nn
class WrappedModel(nn.Module):
  def __init__(self, inner_model):
    super().__init__()
    self.inner = inner_model
  def forward(self, *args, **kwargs):
    return self.inner(*args, **kwargs)

wrapped_model = WrappedModel(WrappedModel(t5))

# say we use LoRA
delta_config = AutoDeltaConfig.from_dict({"delta_type":"lora"})
delta2 = AutoDeltaModel.from_config(delta_config, backbone_model=wrapped_model)
delta2.log()
# >>> root
#       -- inner
#          -- inner
#             ...
#             ... lora_A:[8,1024], lora_B:[1024,8]
delta2.detach()

# use a not default configuration
# say we add lora to the last four layer of the decoder of t5, with lora rank=5
delta_config3 = AutoDeltaConfig.from_dict({"delta_type":"lora", "modified_modules":["[r]decoder.*((20)|(21)|(22)|(23)).*DenseReluDense\.wi"], "lora_r":5})
delta3 = AutoDeltaModel.from_config(delta_config3, backbone_model=wrapped_model)
delta3.log()

Verified Default Configurations

• You can try to use OpenDelta on any backbone models based on PyTorch.

• However, with small chances thatThe interface of the submodules of the backbone model is not supported. Therefore we verified some commonly used models that OpenDelta are sure to support.

• We will keep testing more and more emerging models.

• Pull requests are welcomed when you successfully apply OpenDelta on your own backbone model.

	Lora	Bias Tuning	Adapter Houstbly	Adapter Preffier	Adapter Drop	Adapater Low-Rank	Compactor	Prefix Tuning	Prompt Tuning
T5	✅	✅	✅	✅	✅	✅	✅	✅	✅
GPT-2	✅	✅	✅	✅	✅	✅	✅	✅
BART	✅	✅	✅	✅	✅	✅	✅	✅
DistilBERT	✅	✅	✅	✅	✅	✅	✅	✅
RoBERTa	✅	✅	✅	✅	✅	✅	✅	✅
BERT	✅	✅	✅	✅	✅	✅	✅	✅	✅
T5-3b(parallel)	✅	✅	✅	✅	✅	✅	✅	✅	✅
Deberta-v2	✅	✅	✅	✅	✅	✅	✅
CTRL	✅	✅	✅	✅	✅	✅	✅