PyTorch Implementation for AAAI'21 "Do Response Selection Models Really Know What's Next? Utterance Manipulation Strategies for Multi-turn Response Selection"

Taesun Whang

Last update: Nov 22, 2022

Related tags

Deep Learning UMS-ResSel

Overview

UMS for Multi-turn Response Selection

Implements the model described in the following paper Do Response Selection Models Really Know What's Next? Utterance Manipulation Strategies for Multi-turn Response Selection.

@inproceedings{whang2021ums,
  title={Do Response Selection Models Really Know What's Next? Utterance Manipulation Strategies for Multi-turn Response Selection},
  author={Whang, Taesun and Lee, Dongyub and Oh, Dongsuk and Lee, Chanhee and Han, Kijong and Lee, Dong-hun and Lee, Saebyeok},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2021}
}

This code is reimplemented as a fork of huggingface/transformers and taesunwhang/BERT-ResSel.

Setup and Dependencies

This code is implemented using PyTorch v1.6.0, and provides out of the box support with CUDA 10.1 and CuDNN 7.6.5.

Anaconda / Miniconda is the recommended to set up this codebase.

Anaconda or Miniconda

Clone this repository and create an environment:

git clone https://www.github.com/taesunwhang/UMS-ResSel
conda create -n ums_ressel python=3.7

# activate the environment and install all dependencies
conda activate ums_ressel
cd UMS-ResSel

# https://pytorch.org
pip install torch==1.6.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

Preparing Data and Checkpoints

Pre- and Post-trained Checkpoints

We provide following pre- and post-trained checkpoints.

bert-base (english), bert-base-wwm (chinese)
bert-post (ubuntu, douban, e-commerce)
electra-base (english), electra-base (chinese)
electra-post (ubuntu, douban, e-commerce)

sh scripts/download_pretrained_checkpoints.sh

Data pkls for Fine-tuning (Response Selection)

Original version for each dataset is availble in Ubuntu Corpus V1, Douban Corpus, and E-Commerce Corpus, respectively.

sh scripts/download_datasets.sh

Domain-specific Post-Training

Post-training Creation

Data for post-training BERT

#Ubuntu Corpus V1
sh scripts/create_bert_post_data_creation_ubuntu.sh
#Douban Corpus
sh scripts/create_bert_post_data_creation_douban.sh
#E-commerce Corpus
sh scripts/create_bert_post_data_creation_e-commerce.sh

Data for post-training ELECTRA

sh scripts/download_electra_post_training_pkl.sh

Post-training Examples

BERT+ (e.g., Ubuntu Corpus V1)

python3 main.py --model bert_post_training --task_name ubuntu --data_dir data/ubuntu_corpus_v1 --bert_pretrained bert-base-uncased --bert_checkpoint_path bert-base-uncased-pytorch_model.bin --task_type response_selection --gpu_ids "0" --root_dir /path/to/root_dir --training_type post_training

ELECTRA+ (e.g., Douban Corpus)

python3 main.py --model electra_post_training --task_name douban --data_dir data/electra_post_training --bert_pretrained electra-base-chinese --bert_checkpoint_path electra-base-chinese-pytorch_model.bin --task_type response_selection --gpu_ids "0" --root_dir /path/to/root_dir --training_type post_training

Training Response Selection Models

Model Arguments

BERT-Base

task_name	data_dir	bert_pretrained	bert_checkpoint_path
ubuntu	data/ubuntu_corpus_v1	bert-base-uncased	bert-base-uncased-pytorch_model.bin
douban e-commerce	data/douban data/e-commerce	bert-base-wwm-chinese	bert-base-wwm-chinese_model.bin

BERT-Post

task_name	data_dir	bert_pretrained	bert_checkpoint_path
ubuntu	data/ubuntu_corpus_v1	bert-post-uncased	bert-post-uncased-pytorch_model.pth
douban	data/douban	bert-post-douban	bert-post-douban-pytorch_model.pth
e-commerce	data/e-commerce	bert-post-ecommerce	bert-post-ecommerce-pytorch_model.pth

ELECTRA-Base

task_name	data_dir	bert_pretrained	bert_checkpoint_path
ubuntu	data/ubuntu_corpus_v1	electra-base	electra-base-pytorch_model.bin
douban e-commerce	data/douban data/e-commerce	electra-base-chinese	electra-base-chinese-pytorch_model.bin

ELECTRA-Post

task_name	data_dir	bert_pretrained	bert_checkpoint_path
ubuntu	data/ubuntu_corpus_v1	electra-post	electra-post-pytorch_model.pth
douban	data/douban	electra-post-douban	electra-post-douban-pytorch_model.pth
e-commerce	data/e-commerce	electra-post-ecommerce	electra-post-ecommerce-pytorch_model.pth

Fine-tuning Examples

BERT+ (e.g., Ubuntu Corpus V1)

python3 main.py --model bert_post --task_name ubuntu --data_dir data/ubuntu_corpus_v1 --bert_pretrained bert-post-uncased --bert_checkpoint_path bert-post-uncased-pytorch_model.pth --task_type response_selection --gpu_ids "0" --root_dir /path/to/root_dir

UMS BERT+ (e.g., Douban Corpus)

python3 main.py --model bert_post --task_name douban --data_dir data/douban --bert_pretrained bert-post-douban --bert_checkpoint_path bert-post-douban-pytorch_model.pth --task_type response_selection --gpu_ids "0" --root_dir /path/to/root_dir --multi_task_type "ins,del,srch"

UMS ELECTRA (e.g., E-Commerce)

python3 main.py --model electra_base --task_name e-commerce --data_dir data/e-commerce --bert_pretrained electra-base-chinese --bert_checkpoint_path electra-base-chinese-pytorch_model.bin --task_type response_selection --gpu_ids "0" --root_dir /path/to/root_dir --multi_task_type "ins,del,srch"

Evaluation

To evaluate the model, set --evaluate to /path/to/checkpoints

UMS BERT+ (e.g., Ubuntu Corpus V1)

python3 main.py --model bert_post --task_name ubuntu --data_dir data/ubuntu_corpus_v1 --bert_pretrained bert-post-uncased --bert_checkpoint_path bert-post-uncased-pytorch_model.pth --task_type response_selection --gpu_ids "0" --root_dir /path/to/root_dir --evaluate /path/to/checkpoints --multi_task_type "ins,del,srch"

Performance

We provide model checkpoints of UMS-BERT+, which obtained new state-of-the-art, for each dataset.

Ubuntu	R@1	R@2	R@5
UMS-BERT+	0.875	0.942	0.988

Douban	MAP	MRR	P@1	R@1	R@2	R@5
UMS-BERT+	0.625	0.664	0.499	0.318	0.482	0.858

E-Commerce	R@1	R@2	R@5
UMS-BERT+	0.762	0.905	0.986

Comments

How do you choose checkpoint during training?
Hi, thanks for open source this wonderful work! I have a question about how to choose the checkpoint during training, since I notice that in the config file, the parameter evaluate_data_type is set to test.

ECOMMERCE_PARAMS = defaultdict( evaluate_candidates_num=10, recall_k_list=[1, 2, 5, 10], evaluate_data_type="test", language="chinese", max_utt_len=5, )
opened by FFYYang 5
Recall@1 achieves 100% on the Ubuntu dataset
Hi

I follow the instruction in the README, to be specified

download pre-trained and post-trained checkpoints and dataset by running the two scripts

run fine-tuning through ``python3 main.py --model bert_post --task_name ubuntu --data_dir data/ubuntu_corpus_v1 --bert_pretrained bert-post-uncased --bert_checkpoint_path bert-post-uncased-pytorch_model.pth --task_type response_selection --gpu_ids "0,1,2,3" --root_dir./ '' and then the recall@1 on Ubuntu dataset achieves 100% at the second epoch

Could you please tell me is there anything wrong with my operations?
opened by JuruoMP 5
The difference in evaluation utils between the previous works and yours

Hello, thanks for your wonderful work.

Except for your codes, I also read other released response selection repos, and I found the difference in evaluation utils (only for douban corpus).

In SA-BERT, it can be found that, during calculating the R10@x scores, the previous work count when one positive sample are retrieved. But in your codes, you calculate the average. I am not sure which one is more appropriate, can you explain it to me?

Thank you very much!

opened by gmftbyGMFTBY 3
Maybe a bug in dataset.py?

Thanks for open source your great work! I don't really understand what does this line used for. https://github.com/taesunwhang/UMS-ResSel/blob/master/data/dataset.py#L93

opened by FFYYang 2
Hardware of training BERT post model

Hi, after reading your codes of post train, I realize that you leverage the huggingface BertForPreTraning module. I reconstruct the post train codes, but I find it cost me about 12 hours to train one epoch. My devices are 8 1080 Ti GPUs.

I want to know the hardware configuration during your experiments, and how much time cost for training one epoch in your settings.

opened by gmftbyGMFTBY 2
Maybe wrong grad clipping

https://github.com/taesunwhang/UMS-ResSel/blob/e261f5c7d31bcac92e9ef72a47e7a1727226fc91/post_train/post_training.py#L162

Hi, first of all, thank you for your wonderful work, I am very confused about the gradient clipping position. To the best of my knowledge, the grad clipping should be placed before optimizer.step, or it will never be used.

But I am not sure about this issue. Can you explain this question to me?

opened by gmftbyGMFTBY 2

PyTorch Implementation for AAAI'21 "Do Response Selection Models Really Know What's Next? Utterance Manipulation Strategies for Multi-turn Response Selection"

Related tags

Overview

UMS for Multi-turn Response Selection

Setup and Dependencies

Anaconda or Miniconda

Preparing Data and Checkpoints

Pre- and Post-trained Checkpoints

Data pkls for Fine-tuning (Response Selection)

Domain-specific Post-Training

Post-training Creation

Data for post-training BERT

Data for post-training ELECTRA

Post-training Examples

BERT+ (e.g., Ubuntu Corpus V1)

ELECTRA+ (e.g., Douban Corpus)

Training Response Selection Models

Model Arguments

BERT-Base

BERT-Post

ELECTRA-Base

ELECTRA-Post

Fine-tuning Examples

BERT+ (e.g., Ubuntu Corpus V1)

UMS BERT+ (e.g., Douban Corpus)

UMS ELECTRA (e.g., E-Commerce)

Evaluation

UMS BERT+ (e.g., Ubuntu Corpus V1)

Performance

Comments

How do you choose checkpoint during training?

Recall@1 achieves 100% on the Ubuntu dataset

The difference in evaluation utils between the previous works and yours

Maybe a bug in dataset.py?

Hardware of training BERT post model

Maybe wrong grad clipping

Owner

Taesun Whang

RealFormer-Pytorch Implementation of RealFormer using pytorch

A PyTorch implementation of the paper Mixup: Beyond Empirical Risk Minimization in PyTorch

A pytorch implementation of Pytorch-Sketch-RNN

PyTorch implementation of Advantage async actor-critic Algorithms (A3C) in PyTorch

Pytorch-diffusion - A basic PyTorch implementation of 'Denoising Diffusion Probabilistic Models'

RetinaNet-PyTorch - A RetinaNet Pytorch Implementation on remote sensing images and has the similar mAP result with RetinaNet in MMdetection

RETRO-pytorch - Implementation of RETRO, Deepmind's Retrieval based Attention net, in Pytorch

HashNeRF-pytorch - Pure PyTorch Implementation of NVIDIA paper on Instant Training of Neural Graphics primitives

Generic template to bootstrap your PyTorch project with PyTorch Lightning, Hydra, W&B, and DVC.

A PyTorch Extension: Tools for easy mixed precision and distributed training in Pytorch

Objective of the repository is to learn and build machine learning models using Pytorch. 30DaysofML Using Pytorch

Pretrained SOTA Deep Learning models, callbacks and more for research and production with PyTorch Lightning and PyTorch

Amazon Forest Computer Vision: Satellite Image tagging code using PyTorch / Keras with lots of PyTorch tricks

The Incredible PyTorch: a curated list of tutorials, papers, projects, communities and more relating to PyTorch.

Amazon Forest Computer Vision: Satellite Image tagging code using PyTorch / Keras with lots of PyTorch tricks

A bunch of random PyTorch models using PyTorch's C++ frontend

PyTorch Autoencoders - Implementing a Variational Autoencoder (VAE) Series in Pytorch.

PyTorch-LIT is the Lite Inference Toolkit (LIT) for PyTorch which focuses on easy and fast inference of large models on end-devices.

A general framework for deep learning experiments under PyTorch based on pytorch-lightning