Text Generation by Learning from Demonstrations

Overview

Text Generation by Learning from Demonstrations

The README was last updated on March 7, 2021. The repo is based on fairseq (v0.9.?).

Paper

arXiv

Prerequisites

Per fairseq usage, we need to install this particular modifed version fairseq. The simplest way: pip install --editable ./.

Due to pytorch changes, and given that we're using a slightly older version of fairseq (see below), please use pytorch version <= 1.6.0. However, the GOLD algorithm can be easily implemented on top of the latest fairseq (or most text generation codebases).

Datasets

For downloading CNN/DM and XSum datasets, we follow the instructions here; note that this link does not correspond to the latest fairseq. Our version of the CNN/DM input articles include the prepended "(CNN)" tags. For downloading IWSLT14 De-En dataset, we follow the instructions here. The binary files are provided in our repo, in the directory data-bin. For downloading the particular version of our NQG dataset, we follow the instructions here. The binary files are provided upon request.

Code: experiments on transformer models using fairseq

For reproducibility, the code is based on a April 2020 version of fairseq (based on release v0.9.0). However, it is easy to reimplement the GOLD algorithm in the latest version of fairseq and in another frameworks.

How to implement in the latest version of fairseq?

  • If your GPUs "have large memory", then most of the implementation happens around the criterion code (for question generation, summarization, translation, the py file is ./fairseq/criterions/label_smoothed_cross_entropy.py in the April 2020 version of fairseq). Note that the implementation in this repo uses this approach.
    • "Have large memory": Meaning the GPUs can store pi, pi-tilde, p_MLE at the same time; see Algorithm 1 in the paper. In our experiments (using the same datasets, same batch size, etc.), this would imply that the GPUs have ~24G of memory.
  • If your GPUs cannot fit the above models, then you may need to input p_MLE probabilities as features. This can be done by first saving the probabilities into a text file or pickle file, and then loading them in the load_langpair_dataset function of ./fairseq/tasks/translation.py (or other corresponding files for other tasks).

How to implement in other codebase?

  • See Algorithm 1 in the paper. The majority of the work will happen around the loss computation. We need to have three different models ready when computing losses: (1) pi, the network we're training; (2) pi-tilde, a slightly older version of pi (created to ensure training stability, similar to the periodic synchronization in deep Q-learning; (3) p_MLE, to compute rewards (but this can be pre-loaded in the form of input features, in case the GPU cannot fit the third model).

BART summarization generation fairseq issue

Given that there has been minor bugs with the fairseq BART summarization code (details on original fairseq github), we make the corresponding changes according to the fairseq authors' recommendation. (1) In ./fairseq/sequence_generator.py, see the modification here. (2) In ./fairseq/tasks/fairseq_task.py, see the modification here. (3) In ./fairseq/models/bart/hub_interface.py, see the modification here. The above is already implemented in this repo. But if we're reimplementing the GOLD code in the latest fairseq, we need to beware of this issue (and keep the three modifications in mind).

How to run?

Training

The entry point for training is ./fairseq_cli/train.py. See ./fairseq/options.py for possible flags. For CNN/DM, the script for running GOLD-p is provided in run_cnndm_goldp.sh; the script for running GOLD-s (which often performs better than GOLD-p) is provided in run_cnndm_golds.sh. Some other scripts for other tasks are also provided. For explanations of flags, please refer to ./fairseq/options.py as well as Algorithm 1 in the paper.

Validation

Note that to validate, one possibility is to find the checkpoint that corresponds to highest BLEU/ROUGE-2 score on dev set. We cannot validate according to NLL loss, given that in the paper, we showed that our models achieve higher accuracy but higher perplexity (and NLL loss). Do not use checkpoint_best.pt. IWSLT14 De-En validation is implemented. For summarization, please use run_cnndm_validation.py (similar to run_cnndm_inference.py) as an example to loop through all checkpoints. Then, compute the ROUGE based on run_cnndm_validation_step2.sh (perhaps with small modifications).

Evaluation/inference

For BART evaluation, we use the inference scripts provided in run_cnndm_inference.sh, run_xsum_inference.sh, run_squad_inference.sh. For IWSLT14 De-En inference, the following few lines will do.

python -W ignore [path-to-fairseq_cli/generate.py] data-bin/iwslt14.tokenized.de-en \
    --path [path-to-model-checkpoint.pt] \
    --batch-size 128 --beam 5 --remove-bpe --gen-subset test  > [path-to-save-to-file]

Transformer models

Please ensure the data is processed appropriately before using the models.

MLE model checkpoints

GOLD-s model checkpoints

Not a lot of hyperparameter search was done for the transformer models, so it is likely that more search (on hyperparameters, on architecture) could reach better performance.

Moreover, for summarization models, we use pyrouge+files2rouge to evaluate, based on the fairseq instructions after pyrouge+files2rouge installation. The package files2rouge has a common WordNet-2.0.exc.db error; see this link for the fix.

Citation, authors, and contact

The bibtex entry

Richard Yuanzhe Pang

He He

A collection of loss functions for medical image segmentation

A collection of loss functions for medical image segmentation

Jun 3.1k Jan 03, 2023
DeiT: Data-efficient Image Transformers

DeiT: Data-efficient Image Transformers This repository contains PyTorch evaluation code, training code and pretrained models for DeiT (Data-Efficient

Facebook Research 3.2k Jan 06, 2023
This is an easy python software which allows to sort images with faces by gender and after by age.

Gender-age Classifier This is an easy python software which allows to sort images with faces by gender and after by age. Usage First install Deepface

Claudio Ciccarone 6 Sep 17, 2022
The official repository for Deep Image Matting with Flexible Guidance Input

FGI-Matting The official repository for Deep Image Matting with Flexible Guidance Input. Paper: https://arxiv.org/abs/2110.10898 Requirements easydict

Hang Cheng 51 Nov 10, 2022
Code for "Continuous-Time Meta-Learning with Forward Mode Differentiation" (ICLR 2022)

Continuous-Time Meta-Learning with Forward Mode Differentiation ICLR 2022 (Spotlight) - Installation - Example - Citation This repository contains the

Tristan Deleu 25 Oct 20, 2022
A check for whether the dependency jobs are all green.

alls-green A check for whether the dependency jobs are all green. Why? Do you have more than one job in your GitHub Actions CI/CD workflows setup? Do

Re:actors 33 Jan 03, 2023
This repository is related to an Arabic tutorial, within the tutorial we discuss the common data structure and algorithms and their worst and best case for each, then implement the code using Python.

Data Structure and Algorithms with Python This repository is related to the Arabic tutorial here, within the tutorial we discuss the common data struc

Mohamed Ayman 33 Dec 02, 2022
The Empirical Investigation of Representation Learning for Imitation (EIRLI)

The Empirical Investigation of Representation Learning for Imitation (EIRLI)

Center for Human-Compatible AI 31 Nov 06, 2022
The project page of paper: Architecture disentanglement for deep neural networks [ICCV 2021, oral]

This is the project page for the paper: Architecture Disentanglement for Deep Neural Networks, Jie Hu, Liujuan Cao, Tong Tong, Ye Qixiang, ShengChuan

Jie Hu 15 Aug 30, 2022
Implementation of trRosetta and trDesign for Pytorch, made into a convenient package

trRosetta - Pytorch (wip) Implementation of trRosetta and trDesign for Pytorch, made into a convenient package

Phil Wang 67 Dec 17, 2022
[ICCV2021] Official Pytorch implementation for SDGZSL (Semantics Disentangling for Generalized Zero-Shot Learning)

Semantics Disentangling for Generalized Zero-shot Learning This is the official implementation for paper Zhi Chen, Yadan Luo, Ruihong Qiu, Zi Huang, J

25 Dec 06, 2022
VID-Fusion: Robust Visual-Inertial-Dynamics Odometry for Accurate External Force Estimation

VID-Fusion VID-Fusion: Robust Visual-Inertial-Dynamics Odometry for Accurate External Force Estimation Authors: Ziming Ding , Tiankai Yang, Kunyi Zhan

ZJU FAST Lab 86 Nov 18, 2022
Official Implementation of DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation

DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation [Arxiv] [Paper] As acquiring pixel-wise an

Lukas Hoyer 305 Dec 29, 2022
DecoupledNet is semantic segmentation system which using heterogeneous annotations

DecoupledNet: Decoupled Deep Neural Network for Semi-supervised Semantic Segmentation Created by Seunghoon Hong, Hyeonwoo Noh and Bohyung Han at POSTE

Hyeonwoo Noh 74 Sep 22, 2021
Illuminated3D This project participates in the Nasa Space Apps Challenge 2021.

Illuminated3D This project participates in the Nasa Space Apps Challenge 2021.

Eleftheriadis Emmanouil 1 Oct 09, 2021
Code for the CVPR 2021 paper: Understanding Failures of Deep Networks via Robust Feature Extraction

Welcome to Barlow Barlow is a tool for identifying the failure modes for a given neural network. To achieve this, Barlow first creates a group of imag

Sahil Singla 33 Dec 05, 2022
Official implementation of the paper "Light Field Networks: Neural Scene Representations with Single-Evaluation Rendering"

Light Field Networks Project Page | Paper | Data | Pretrained Models Vincent Sitzmann*, Semon Rezchikov*, William Freeman, Joshua Tenenbaum, Frédo Dur

Vincent Sitzmann 130 Dec 29, 2022
Tiny Kinetics-400 for test

Kinetics-400迷你数据集 English | 简体中文 该数据集旨在解决的问题:参照Kinetics-400数据格式,训练基于自己数据的视频理解模型。 数据集介绍 Kinetics-400是视频领域benchmark常用数据集,详细介绍可以参考其官方网站Kinetics。整个数据集包含40

38 Jan 06, 2023
BEAMetrics: Benchmark to Evaluate Automatic Metrics in Natural Language Generation

BEAMetrics: Benchmark to Evaluate Automatic Metrics in Natural Language Generation Installing The Dependencies $ conda create --name beametrics python

7 Jul 04, 2022
Semi-supervised Implicit Scene Completion from Sparse LiDAR

Semi-supervised Implicit Scene Completion from Sparse LiDAR Paper Created by Pengfei Li, Yongliang Shi, Tianyu Liu, Hao Zhao, Guyue Zhou and YA-QIN ZH

114 Nov 30, 2022