When BERT Plays the Lottery, All Tickets Are Winning

Overview

When BERT Plays the Lottery, All Tickets Are Winning

Large Transformer-based models were shown to be reducible to a smaller number of self-attention heads and layers. We consider this phenomenon from the perspective of the lottery ticket hypothesis, using both structured and magnitude pruning. For fine-tuned BERT, we show that (a) it is possible to find subnetworks achieving performance that is comparable with that of the full model, and (b) similarly-sized subnetworks sampled from the rest of the model perform worse. Strikingly, with structured pruning even the worst possible subnetworks remain highly trainable, indicating that most pre-trained BERT weights are potentially useful. We also study the "good" subnetworks to see if their success can be attributed to superior linguistic knowledge, but find them unstable, and not explained by meaningful self-attention patterns.

Environment

Install the requirements in your python 3.7.7 virtual environment.

pip install -r requirements.txt

These experiments were done on multi-gpu environment, were some experiments, benchmarks were run parallel. So some changes to the bash scripts to make it work for your environment.

Dataset

  1. Download the GLUE dataset using data/download_glue.py and data/download_mnli_data.py. Follow the instructions in data/download_glue.py docstring for MRPC.
  2. All data for the tasks should be organized in data/glue/task_name/ structure.
  3. Extract the attention pattern classification labelled data.
    cd data
    tar -xvf head_classification_data.tar.gz

Training, Masking, and Evaluation

Switch cwd to src (cd src) as many paths are relative from that directory.

  1. Fine-tune the BERT on GLUE tasks
./train.sh
  1. Obtain the masks
./find_masks.sh
  1. Train models with the masks applied in good, random and bad settings.
./train_with_masks.sh
  1. Evaluate the trained models
./evaluate.sh

Note: These experiments were run through course of time and now stiched together into single scripts. So it might be better to run the training and evaluation commands in them one by one.

  1. Train the CNN classifier on attention patterns normed and raw.
python classify_attention_patterns.py
python classify_normed_patterns.py

These only train the classifier.

Evaluation Analysis and Final Results

These are primarily done in jupyter notebooks in experiment_analysis directory. There are many experimental notebooks there. Here are the important ones used to generate results included in the paper.

  1. Importance pruning Heatmaps. Ignore the final "train_subset" and "hans" settings.
  2. Magnitude pruning Heatmap
  3. Overlap of surviving components
  4. Generate the random baseline
  5. Attention Classification Patterns
  6. Evaluation Result Comparisons and table
  7. Statistics on mask correlation across seeds
Owner
Sai
Machine Learning Researcher - NLP
Sai
Iowa Project - My second project done at General Assembly, focused on feature engineering and understanding Linear Regression as a concept

Project 2 - Ames Housing Data and Kaggle Challenge PROBLEM STATEMENT Inferring or Predicting? What's more valuable for a housing model? When creating

Adam Muhammad Klesc 1 Jan 03, 2022
Unsupervised Image to Image Translation with Generative Adversarial Networks

Unsupervised Image to Image Translation with Generative Adversarial Networks Paper: Unsupervised Image to Image Translation with Generative Adversaria

Hao 71 Oct 30, 2022
Lightweight tool to perform MITM attack on local network

ARPSpy - A lightweight tool to perform MITM attack Using many library to perform ARP Spoof and auto-sniffing HTTP packet containing credential. (Never

MinhItachi 8 Aug 28, 2022
Determined: Deep Learning Training Platform

Determined: Deep Learning Training Platform Determined is an open-source deep learning training platform that makes building models fast and easy. Det

Determined AI 2k Dec 31, 2022
load .txt to train YOLOX, same as Yolo others

YOLOX train your data you need generate data.txt like follow format (per line- one image). prepare one data.txt like this: img_path1 x1,y1,x2,y2,clas

LiMingf 18 Aug 18, 2022
An image base contains 490 images for learning (400 cars and 90 boats), and another 21 images for testingAn image base contains 490 images for learning (400 cars and 90 boats), and another 21 images for testing

SVM Données Une base d’images contient 490 images pour l’apprentissage (400 voitures et 90 bateaux), et encore 21 images pour fait des tests. Prétrait

Achraf Rahouti 3 Nov 30, 2021
Cache Requests in Deta Bases and Echo them with Deta Micros

Deta Echo Cache Leverage the awesome Deta Micros and Deta Base to cache requests and echo them as needed. Stop worrying about slow public APIs or agre

Gingerbreadfork 8 Dec 07, 2021
ToFFi - Toolbox for Frequency-based Fingerprinting of Brain Signals

ToFFi Toolbox This repository contains "before peer review" version of the software related to the preprint of the publication ToFFi - Toolbox for Fre

4 Aug 31, 2022
The ARCA23K baseline system

ARCA23K Baseline System This is the source code for the baseline system associated with the ARCA23K dataset. Details about ARCA23K and the baseline sy

4 Jul 02, 2022
[ICCV'2021] "SSH: A Self-Supervised Framework for Image Harmonization", Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, Zhangyang Wang

SSH: A Self-Supervised Framework for Image Harmonization (ICCV 2021) code for SSH Representative Examples Main Pipeline RealHM DataSet Google Drive Pr

VITA 86 Dec 02, 2022
A simple and lightweight genetic algorithm for optimization of any machine learning model

geneticml This package contains a simple and lightweight genetic algorithm for optimization of any machine learning model. Installation Use pip to ins

Allan Barcelos 8 Aug 10, 2022
Kaggle Ultrasound Nerve Segmentation competition [Keras]

Ultrasound nerve segmentation using Keras (1.0.7) Kaggle Ultrasound Nerve Segmentation competition [Keras] #Install (Ubuntu {14,16}, GPU) cuDNN requir

179 Dec 28, 2022
This project is used for the paper Differentiable Programming of Isometric Tensor Network

This project is used for the paper "Differentiable Programming of Isometric Tensor Network". (arXiv:2110.03898)

Chenhua Geng 15 Dec 13, 2022
meProp: Sparsified Back Propagation for Accelerated Deep Learning

meProp The codes were used for the paper meProp: Sparsified Back Propagation for Accelerated Deep Learning with Reduced Overfitting (ICML 2017) [pdf]

LancoPKU 107 Nov 18, 2022
Implementation of the Paper: "Parameterized Hypercomplex Graph Neural Networks for Graph Classification" by Tuan Le, Marco Bertolini, Frank Noé and Djork-Arné Clevert

Parameterized Hypercomplex Graph Neural Networks (PHC-GNNs) PHC-GNNs (Le et al., 2021): https://arxiv.org/abs/2103.16584 PHM Linear Layer Illustration

Bayer AG 26 Aug 11, 2022
Joint project of the duo Hacker Ninjas

Project Smoothie Společný projekt dua Hacker Ninjas. První pokus o hříčku po třech týdnech učení se programování. Jakub Kolář e:\

Jakub Kolář 2 Jan 07, 2022
NeurIPS 2021 Datasets and Benchmarks Track

AP-10K: A Benchmark for Animal Pose Estimation in the Wild Introduction | Updates | Overview | Download | Training Code | Key Questions | License Intr

AP-10K 82 Dec 11, 2022
A pytorch-based deep learning framework for multi-modal 2D/3D medical image segmentation

A 3D multi-modal medical image segmentation library in PyTorch We strongly believe in open and reproducible deep learning research. Our goal is to imp

Adaloglou Nikolas 1.2k Dec 27, 2022
Physics-Aware Training (PAT) is a method to train real physical systems with backpropagation.

Physics-Aware Training (PAT) is a method to train real physical systems with backpropagation. It was introduced in Wright, Logan G. & Onodera, Tatsuhiro et al. (2021)1 to train Physical Neural Networ

McMahon Lab 230 Jan 05, 2023
Implementation of Wasserstein adversarial attacks.

Stronger and Faster Wasserstein Adversarial Attacks Code for Stronger and Faster Wasserstein Adversarial Attacks, appeared in ICML 2020. This reposito

21 Oct 06, 2022