This is the implementation of the paper AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning (https://arxiv.org/abs/2205.12410).
APACHE-2.0 License
This is the implementation of the paper AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning.
Our experiments on the GLUE benchmark are run on 16 NVIDIA Tesla V100 GPU. The results may vary due to different GPU models, drivers, CUDA SDK versions, floating-point precisions, and random seeds.
We release all copies of Adapter weights for users' Adapter aggregation study.
Dataset | BERT base 110M | RoBERTa large 355M | |
---|---|---|---|
MNLI | 8.5 MB | 11.7 MB | |
SST2 | 8.5 MB | 11.7 MB | |
MRPC | 8.5 MB | 11.7 MB | |
CoLA | 8.5 MB | 11.7 MB | |
QNLI | 8.5 MB | 11.7 MB | |
QQP | 8.5 MB | 11.7 MB | |
RTE | 8.5 MB | 11.7 MB | |
STSB | 8.5 MB | 11.7 MB |
conda env create -f environment.yml
pip install -e .
We also provide the shell scripts for bert-base and roberta-large.
export num_gpus=1
export PYTHONHASHSEED=0
task_name=mnli
model=roberta-large
export output_dir="./models/${model}/${task_name}"
python -m torch.distributed.launch --nproc_per_node=$num_gpus \
examples/text-classification/run_glue.py \
--model_name_or_path $model \
--task_name $task_name \
--do_train \
--do_eval \
--max_seq_length 128 \
--per_device_train_batch_size 64 \
--per_device_eval_batch_size 32 \
--learning_rate 3e-4 \
--num_train_epochs 20 \
--output_dir $output_dir/model \
--overwrite_output_dir \
--logging_steps 1000 \
--logging_dir $output_dir/log \
--evaluation_strategy epoch \
--save_strategy epoch \
--warmup_ratio 0.06 \
--apply_expert_soup \
--adapter_size 16 \
--num_experts 4 \
--seed 0 \
--inference_level 3 \
--weight_decay 0.1 \
--sharing_up 1 \
--sharing_down 0 \
--use_consistency_loss 1
Most arguments are inherited from transformers and are easy to understand. We further explain some of the AdaMix's arguments:
inference_level
: There are two suggested modes
1
: Random Routing3
: Averaging the weights of Adapters for routing (used in AdaMix)num_experts
: Number of Adapters in AdaMix
use_consistency_loss
: Two modes.
0
: No consistency loss1
: Use consistency losssharing_up
: There are two modes. (sharing_down is same)
0
: No weight sharing1
: Sharing Project-up layer weights in AdapterCreate checkpoints directory and download checkpoints of corresponding tasks under the directory. Use MNLI as an example. Use your checkpoint path in expert_soup_path argument.
export num_gpus=1
export PYTHONHASHSEED=0
task_name=mnli
model=roberta-large
export output_dir="./models/${model}/${task_name}"
python -m torch.distributed.launch --nproc_per_node=$num_gpus \
examples/text-classification/run_glue.py \
--model_name_or_path $model \
--task_name $task_name \
--do_eval \
--expert_soup_path ./checkpoints/pytorch_model_${task_name}_expert_soup.bin \
--max_seq_length 128 \
--per_device_train_batch_size 64 \
--per_device_eval_batch_size 32 \
--learning_rate 3e-4 \
--num_train_epochs 20 \
--output_dir $output_dir/model \
--overwrite_output_dir \
--logging_steps 1000 \
--logging_dir $output_dir/log \
--evaluation_strategy epoch \
--save_strategy epoch \
--warmup_ratio 0.06 \
--apply_expert_soup \
--adapter_size 16 \
--num_experts 4 \
--seed 0 \
--inference_level 3 \
--weight_decay 0.1 \
--sharing_up 1 \
--sharing_down 0 \
--use_consistency_loss 1
The implementation is based on https://github.com/huggingface/transformers We also used some code from: https://github.com/microsoft/LoRA
For personal communication related to this package, please contact Yaqing Wang ([email protected]), Sahaj Agarwal ([email protected]), Subhabrata (Subho) Mukherjee ([email protected]) or Xiaodong Liu ([email protected]).
@article{wang2022adamix,
title={AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning},
author={Wang, Yaqing and Agarwal, Sahaj and Mukherjee, Subhabrata and Liu, Xiaodong and Gao, Jing and Awadallah, Ahmed Hassan and Gao, Jianfeng},
journal={arXiv preprint arXiv:2205.12410},
year={2022}
}