Training vision models with full-batch gradient descent and regularization

Last update: Jan 06, 2023

Related tags

Overview

Stochastic Training is Not Necessary for Generalization -- Training competitive vision models without stochasticity

This repository implements training routines to train vision models for classification on CIFAR-10 without SGD as described in our publication https://arxiv.org/abs/2109.14119.

Abstract:

It is widely believed that the implicit regularization of SGD is fundamental to the impressive generalization behavior we observe in neural networks. In this work, we demonstrate that non-stochastic full-batch training can achieve comparably strong performance to SGD on CIFAR-10 using modern architectures. To this end, we show that the implicit regularization of SGD can be completely replaced with explicit regularization. Our observations indicate that the perceived difficulty of full-batch training is largely the result of its optimization properties and the disproportionate time and effort spent by the ML community tuning optimizers and hyperparameters for small-batch training.

Requirements

PyTorch 1.9.*
Hydra 1.* (via pip install hydra-core --upgrade)
python-lmdb (only for N x CIFAR experiments.)

Pytorch 1.9.* is used for multithreaded persistent workers, _for_each_, functionality, and torch.inference_mode, all of which have to be replaced when using earlier versions.

Training Runs

Scripts for training runs can be found in train.sh

Guide

The main script for fullbatch training should be train_with_gradient_descent.py. This scripts also runs the stochastic gradient descent sanity check with the same codebase. A crucial flag to distinguish between both settings is hyp.train_stochastic=False. The gradient regularization is activated by setting hyp.grad_reg.block_strength=0.5. Have a look at the various folders under config to see all options. The script crunch_loss_landscape.py can measure the loss landscape of checkpointed models which can be used later for visualization.

If you are interesting in extracting components from this repository, you can use the gradient regularization separately, which can be found under fullbatch/models/modules.py in the class GradRegularizer. This regularizer can be instantiated like so:

gradreg = GradRegularizer(model, optimizer, loss_fn, block_strength=0.5, acc_strength=0.0, eps=1e-2,
                          implementation='finite_diff', mixed_precision=False)

and then used during training to modify a list of gradients (such as from torch.autograd.grad) in-place:

grads = gradreg(grads, inputs, labels, pre_grads=None)

Contact

Just open an issue here on github or send us an email if you have any questions.

You might also like...

[CVPR 2021] "The Lottery Tickets Hypothesis for Supervised and Self-supervised Pre-training in Computer Vision Models" Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Michael Carbin, Zhangyang Wang

The Lottery Tickets Hypothesis for Supervised and Self-supervised Pre-training in Computer Vision Models Codes for this paper The Lottery Tickets Hypo

59 Dec 28, 2022

Stochastic Downsampling for Cost-Adjustable Inference and Improved Regularization in Convolutional Networks

Stochastic Downsampling for Cost-Adjustable Inference and Improved Regularization in Convolutional Networks (SDPoint) This repository contains the cod

17 Jul 4, 2022

[ICLR 2021] Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization

Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization Kaidi Cao, Yining Chen, Junwei Lu, Nikos Arechiga, Adrien Gaidon, Tengyu Ma

29 Oct 20, 2022

PyTorch framework, for reproducing experiments from the paper Implicit Regularization in Hierarchical Tensor Factorization and Deep Convolutional Neural Networks

Implicit Regularization in Hierarchical Tensor Factorization and Deep Convolutional Neural Networks. Code, based on the PyTorch framework, for reprodu

3 Dec 27, 2022

Consistency Regularization for Adversarial Robustness

Consistency Regularization for Adversarial Robustness Official PyTorch implementation of Consistency Regularization for Adversarial Robustness by Jiho

40 Dec 17, 2022

Code for CoMatch: Semi-supervised Learning with Contrastive Graph Regularization

CoMatch: Semi-supervised Learning with Contrastive Graph Regularization (Salesforce Research) This is a PyTorch implementation of the CoMatch paper [B

107 Dec 14, 2022

The code release of paper 'Domain Generalization for Medical Imaging Classification with Linear-Dependency Regularization' NIPS 2020.

Domain Generalization for Medical Imaging Classification with Linear Dependency Regularization The code release of paper 'Domain Generalization for Me

56 Dec 28, 2022

PyTorch implementation of Self-supervised Contrastive Regularization for DG (SelfReg)

SelfReg PyTorch official implementation of Self-supervised Contrastive Regularization for Domain Generalization (SelfReg, https://arxiv.org/abs/2104.0

64 Dec 16, 2022

Code for ACL2021 paper Consistency Regularization for Cross-Lingual Fine-Tuning.

xTune Code for ACL2021 paper Consistency Regularization for Cross-Lingual Fine-Tuning. Environment DockerFile: dancingsoul/pytorch:xTune Install the f

42 Dec 9, 2022

Training vision models with full-batch gradient descent and regularization

Related tags

Overview

Stochastic Training is Not Necessary for Generalization -- Training competitive vision models without stochasticity

Abstract:

Requirements

Training Runs

Guide

Contact

You might also like...

[CVPR 2021] "The Lottery Tickets Hypothesis for Supervised and Self-supervised Pre-training in Computer Vision Models" Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Michael Carbin, Zhangyang Wang

Stochastic Downsampling for Cost-Adjustable Inference and Improved Regularization in Convolutional Networks

[ICLR 2021] Heteroskedastic and Imbalanced Deep Learning with Adaptive Regularization

PyTorch framework, for reproducing experiments from the paper Implicit Regularization in Hierarchical Tensor Factorization and Deep Convolutional Neural Networks

Consistency Regularization for Adversarial Robustness

Code for CoMatch: Semi-supervised Learning with Contrastive Graph Regularization

The code release of paper 'Domain Generalization for Medical Imaging Classification with Linear-Dependency Regularization' NIPS 2020.

PyTorch implementation of Self-supervised Contrastive Regularization for DG (SelfReg)

Code for ACL2021 paper Consistency Regularization for Cross-Lingual Fine-Tuning.

Releases(v1)

v1(Dec 15, 2021)

Owner

Jonas Geiping

Hyperbolic Procrustes Analysis Using Riemannian Geometry

3D-Reconstruction 基于深度学习方法的单目多视图三维重建

TransMVSNet: Global Context-aware Multi-view Stereo Network with Transformers.

A Probabilistic End-To-End Task-Oriented Dialog Model with Latent Belief States towards Semi-Supervised Learning

[CVPR2021] The source code for our paper 《Removing the Background by Adding the Background: Towards Background Robust Self-supervised Video Representation Learning》.

Train Yolov4 using NBX-Jobs

Non-Metric Space Library (NMSLIB): An efficient similarity search library and a toolkit for evaluation of k-NN methods for generic non-metric spaces.

Demonstration of the Model Training as a CI/CD System in Vertex AI

Repo for paper "Dynamic Placement of Rapidly Deployable Mobile Sensor Robots Using Machine Learning and Expected Value of Information"

Learning 3D Part Assembly from a Single Image

A Benchmark For Measuring Systematic Generalization of Multi-Hierarchical Reasoning

Using Streamlit to host a multi-page tool with model specs and classification metrics, while also accepting user input values for prediction.

Unofficial PyTorch implementation of Masked Autoencoders Are Scalable Vision Learners

Repository of continual learning papers

Re-implementation of the vector capsule with dynamic routing

Official PyTorch implementation of Retrieve in Style: Unsupervised Facial Feature Transfer and Retrieval.

Source code for Fixed-Point GAN for Cloud Detection

Official PyTorch Implementation of Mask-aware IoU and maYOLACT Detector [BMVC2021]

Official implementation of our paper "Learning to Bootstrap for Combating Label Noise"

Unofficial Implementation of Oboe (SIGCOMM'18').