EMNLP 2021: Single-dataset Experts for Multi-dataset Question-Answering

Related tags

Deep LearningMADE
Overview

MADE (Multi-Adapter Dataset Experts)

This repository contains the implementation of MADE (Multi-adapter dataset experts), which is described in the paper Single-dataset Experts for Multi-dataset Question Answering.

MADE combines a shared Transformer with a collection of adapters that are specialized to different reading comprehension datasets. See our paper for details.

Quick links

Requirements

The code uses Python 3.8, PyTorch, and the adapter-transformers library. Install the requirements with:

pip install -r requirements.txt

Download the data

You can download the datasets used in the paper from the repository for the MRQA 2019 shared task.

The datasets should be stored in directories ending with train or dev. For example, download the in-domain training datasets to a directory called data/train/ and download the in-domain development datasets to data/dev/.

For zero-shot and few-shot experiments, download the MRQA out-of-domain development datasets to a separate directory and split them into training and development splits using scripts/split_datasets.py. For example, download the datasets to data/transfer/ and run

ls data/transfer/* -1 | xargs -l python scripts/split_datasets.py

Use the default random seed (13) to replicate the splits used in the paper.

Download the trained models

The trained models are stored on the HuggingFace model hub at this URL: https://huggingface.co/princeton-nlp/MADE. All of the models are based on the RoBERTa-base model. They are:

To download just the MADE Transformer and adapters:

mkdir made_transformer
wget https://huggingface.co/princeton-nlp/MADE/resolve/main/made_transformer/model.pt -O made_transformer/model.pt

mkdir made_tuned_adapters
for d in SQuAD HotpotQA TriviaQA SearchQA NewsQA NaturalQuestions; do
  mkdir "made_tuned_adapters/${d}"
  wget "https://huggingface.co/princeton-nlp/MADE/resolve/main/made_tuned_adapters/${d}/model.pt" -O "made_tuned_adapters/${d}/model.pt"
done;

You can download all of the models at once by cloning the repository (first installing Git LFS):

git lfs install
git clone https://huggingface.co/princeton-nlp/MADE
mv MADE models

Run the model

The scripts in scripts/train/ and scripts/transfer/ provide examples of how to run the code. For more details, see the descriptions of the command line flags in run.py.

Train

You can use the scripts in scripts/train/ to train models on the MRQA datasets. For example, to train MADE:

./scripts/train/made_training.sh

And to tune the MADE adapters separately on individual datasets:

for d in SQuAD HotpotQA TriviaQA SearchQA NewsQA NaturalQuestions; do
  ./scripts/train/made_adapter_tuning.sh $d
done;

See run.py for details about the command line arguments.

Evaluate

A single fine-tuned model:

python run.py \
    --eval_on BioASQ DROP DuoRC RACE RelationExtraction TextbookQA \
    --load_from multi_dataset_ft \
    --output_dir output/zero_shot/multi_dataset_ft

An individual MADE adapter (e.g. SQuAD):

python run.py \
    --eval_on BioASQ DROP DuoRC RACE RelationExtraction TextbookQA \
    --load_from made_transformer \
    --load_adapters_from made_tuned_adapters \
    --adapter \
    --adapter_name SQuAD \
    --output_dir output/zero_shot/made_tuned_adapters/SQuAD

An individual single-dataset adapter (e.g. SQuAD):

python run.py \
    --eval_on BioASQ DROP DuoRC RACE RelationExtraction TextbookQA \
    --load_adapters_from single_dataset_adapters/ \
    --adapter \
    --adapter_name SQuAD \
    --output_dir output/zero_shot/single_dataset_adapters/SQuAD

An ensemble of MADE adapters. This will run a forward pass through every adapter in parallel.

python run.py \
    --eval_on BioASQ DROP DuoRC RACE RelationExtraction TextbookQA \
    --load_from made_transformer \
    --load_adapters_from made_tuned_adapters \
    --adapter_names SQuAD HotpotQA TriviaQA SearchQA NewsQA NaturalQuestions \
    --made \
    --parallel_adapters  \
    --output_dir output/zero_shot/made_ensemble

Averaging the parameters of the MADE adapters:

python run.py \
    --eval_on BioASQ DROP DuoRC RACE RelationExtraction TextbookQA \
    --load_from made_transformer \
    --load_adapters_from made_tuned_adapters \
    --adapter_names SQuAD HotpotQA TriviaQA SearchQA NewsQA NaturalQuestions \
    --adapter \
    --average_adapters  \
    --output_dir output/zero_shot/made_avg

Running UnifiedQA:

python run.py \
    --eval_on BioASQ DROP DuoRC RACE RelationExtraction TextbookQA \
    --seq2seq \
    --model_name_or_path allenai/unifiedqa-t5-base \
    --output_dir output/zero_shot/unifiedqa

Transfer

The scripts in scripts/transfer/ provide examples of how to run the few-shot transfer learning experiments described in the paper. For example, the following command will repeat for three random seeds: (1) sample 64 training examples from BioASQ, (2) calculate the zero-shot loss of all the MADE adapters on the training examples, (3) average the adapter parameters in proportion to zero-shot loss, (4) hold out 32 training examples for validation data, (5) train the adapter until performance stops improving on the 32 validation examples, and (6) evaluate the adapter on the full development set.

python run.py \
    --train_on BioASQ \
    --adapter_names SQuAD HotpotQA TriviaQA NewsQA SearchQA NaturalQuestions \
    --made \
    --parallel_made \
    --weighted_average_before_training \
    --adapter_learning_rate 1e-5 \
    --steps 200 \
    --patience 10 \
    --eval_before_training \
    --full_eval_after_training \
    --max_train_examples 64 \
    --few_shot \
    --criterion "loss" \
    --negative_examples \
    --save \
    --seeds 7 19 29 \
    --load_from "made_transformer" \
    --load_adapters_from "made_tuned_adapters" \
    --name "transfer/made_preaverage/BioASQ/64"

Bugs or questions?

If you have any questions related to the code or the paper, feel free to email Dan Friedman ([email protected]). If you encounter any problems when using the code, or want to report a bug, you can open an issue. Please try to specify the problem with details so we can help you better and quicker!

Citation

@inproceedings{friedman2021single,
   title={Single-dataset Experts for Multi-dataset QA},
   author={Friedman, Dan and Dodge, Ben and Chen, Danqi},
   booktitle={Empirical Methods in Natural Language Processing (EMNLP)},
   year={2021}
}
Owner
Princeton Natural Language Processing
Princeton Natural Language Processing
Implementation of Shape and Electrostatic similarity metric in deepFMPO.

DeepFMPO v3D Code accompanying the paper "On the value of using 3D-shape and electrostatic similarities in deep generative methods". The paper can be

34 Nov 28, 2022
Code for the TIP 2021 Paper "Salient Object Detection with Purificatory Mechanism and Structural Similarity Loss"

PurNet Project for the TIP 2021 Paper "Salient Object Detection with Purificatory Mechanism and Structural Similarity Loss" Abstract Image-based salie

Jinming Su 4 Aug 25, 2022
TensorFlow GNN is a library to build Graph Neural Networks on the TensorFlow platform.

TensorFlow GNN This is an early (alpha) release to get community feedback. It's under active development and we may break API compatibility in the fut

889 Dec 30, 2022
MultiTaskLearning - Multi Task Learning for 3D segmentation

Multi Task Learning for 3D segmentation Perception stack of an Autonomous Drivin

2 Sep 22, 2022
An architecture that makes any doodle realistic, in any specified style, using VQGAN, CLIP and some basic embedding arithmetics.

Sketch Simulator An architecture that makes any doodle realistic, in any specified style, using VQGAN, CLIP and some basic embedding arithmetics. See

12 Dec 18, 2022
Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes

Neural Scene Flow Fields PyTorch implementation of paper "Neural Scene Flow Fields for Space-Time View Synthesis of Dynamic Scenes", CVPR 2021 [Projec

Zhengqi Li 583 Dec 30, 2022
Progressive Growing of GANs for Improved Quality, Stability, and Variation

Progressive Growing of GANs for Improved Quality, Stability, and Variation — Official TensorFlow implementation of the ICLR 2018 paper Tero Karras (NV

Tero Karras 5.9k Jan 05, 2023
“英特尔创新大师杯”深度学习挑战赛 赛道3:CCKS2021中文NLP地址相关性任务

基于 bert4keras 的一个baseline 不作任何 数据trick 单模 线上 最高可到 0.7891 # 基础 版 train.py 0.7769 # transformer 各层 cls concat 明神的trick https://xv44586.git

孙永松 7 Dec 28, 2021
A large-scale database for graph representation learning

A large-scale database for graph representation learning

Scott Freitas 29 Nov 25, 2022
Learning to Estimate Hidden Motions with Global Motion Aggregation

Learning to Estimate Hidden Motions with Global Motion Aggregation (GMA) This repository contains the source code for our paper: Learning to Estimate

Shihao Jiang (Zac) 221 Dec 18, 2022
😇A pyTorch implementation of the DeepMoji model: state-of-the-art deep learning model for analyzing sentiment, emotion, sarcasm etc

------ Update September 2018 ------ It's been a year since TorchMoji and DeepMoji were released. We're trying to understand how it's being used such t

Hugging Face 865 Dec 24, 2022
Code for "FPS-Net: A convolutional fusion network for large-scale LiDAR point cloud segmentation".

FPS-Net Code for "FPS-Net: A convolutional fusion network for large-scale LiDAR point cloud segmentation", accepted by ISPRS journal of Photogrammetry

15 Nov 30, 2022
Code for "Share With Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency" paper

UNICORN 🦄 Webpage | Paper | BibTex PyTorch implementation of "Share With Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency" pap

118 Jan 06, 2023
How Effective is Incongruity? Implications for Code-mix Sarcasm Detection.

Code for the paper: How Effective is Incongruity? Implications for Code-mix Sarcasm Detection - ICON ACL 2021

2 Jun 05, 2022
Code for "Learning Graph Cellular Automata"

Learning Graph Cellular Automata This code implements the experiments from the NeurIPS 2021 paper: "Learning Graph Cellular Automata" Daniele Grattaro

Daniele Grattarola 37 Oct 26, 2022
A self-supervised learning framework for audio-visual speech

AV-HuBERT (Audio-Visual Hidden Unit BERT) Learning Audio-Visual Speech Representation by Masked Multimodal Cluster Prediction Robust Self-Supervised A

Meta Research 431 Jan 07, 2023
Sample and Computation Redistribution for Efficient Face Detection

Introduction SCRFD is an efficient high accuracy face detection approach which initially described in Arxiv. Performance Precision, flops and infer ti

Sajjad Aemmi 13 Mar 05, 2022
PyTorch CZSL framework containing GQA, the open-world setting, and the CGE and CompCos methods.

Compositional Zero-Shot Learning This is the official PyTorch code of the CVPR 2021 works Learning Graph Embeddings for Compositional Zero-shot Learni

EML Tübingen 70 Dec 27, 2022
A JAX implementation of Broaden Your Views for Self-Supervised Video Learning, or BraVe for short.

BraVe This is a JAX implementation of Broaden Your Views for Self-Supervised Video Learning, or BraVe for short. The model provided in this package wa

DeepMind 44 Nov 20, 2022
Net2net - Network-to-Network Translation with Conditional Invertible Neural Networks

Net2Net Code accompanying the NeurIPS 2020 oral paper Network-to-Network Translation with Conditional Invertible Neural Networks Robin Rombach*, Patri

CompVis Heidelberg 206 Dec 20, 2022