A Downsampled Variant of ImageNet as an Alternative to the CIFAR datasets

Abstract

The original ImageNet dataset is a popular large-scale benchmark for training Deep Neural Networks. Since the cost of performing experiments (e.g, algorithm design, architecture search, and hyperparameter tuning) on the original dataset might be prohibitive, we propose to consider a downsampled version of ImageNet. In contrast to the CIFAR datasets and earlier downsampled versions of ImageNet, our proposed ImageNet32$\times$32 (and its variants ImageNet64$\times$64 and ImageNet16$\times$16) contains exactly the same number of classes and images as ImageNet, with the only difference that the images are downsampled to 32$\times$32 pixels per image (64$\times$64 and 16$\times$16 pixels for the variants, respectively). Experiments on these downsampled variants are dramatically faster than on the original ImageNet and the characteristics of the downsampled datasets with respect to optimal hyperparameters appear to remain similar. The proposed datasets and scripts to reproduce our results are available at http://image-net.org/download-images and https://github.com/PatrykChrabaszcz/Imagenet32_Scripts

Cited by in corpus (92)

• A Survey on Evolutionary Neural Architecture Search
• Ensemble Distillation for Robust Model Fusion in Federated Learning
• FedALA: Adaptive Local Aggregation for Personalized Federated Learning
• NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size
• Scaling Laws for Autoregressive Generative Modeling
• AutoML-Zero: Evolving Machine Learning Algorithms From Scratch
• FedCP: Separating Feature Information for Personalized Federated Learning via Conditional Policy
• Automatic Perturbation Analysis for Scalable Certified Robustness and Beyond
• LeViT: a Vision Transformer in ConvNet's Clothing for Faster Inference
• NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search
• Deep Feature Space: A Geometrical Perspective
• Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective
• Semantic Segmentation with Labeling Uncertainty and Class Imbalance
• Neural Architecture Search without Training
• Why is the Mahalanobis Distance Effective for Anomaly Detection?
• Red Alarm for Pre-trained Models: Universal Vulnerability to Neuron-Level Backdoor Attacks
• A framework for the extraction of Deep Neural Networks by leveraging public data
• PatchUp: A Feature-Space Block-Level Regularization Technique for Convolutional Neural Networks
• Does Unsupervised Architecture Representation Learning Help Neural Architecture Search?
• Quasi-Global Momentum: Accelerating Decentralized Deep Learning on Heterogeneous Data
• Evaluating Efficient Performance Estimators of Neural Architectures
• Randomized Smoothing of All Shapes and Sizes
• Lossless Image Compression through Super-Resolution
• Defining Benchmarks for Continual Few-Shot Learning
• FlexConv: Continuous Kernel Convolutions with Differentiable Kernel Sizes
• Extended T: Learning with Mixed Closed-set and Open-set Noisy Labels
• Two Sides of the Same Coin: White-box and Black-box Attacks for Transfer Learning
• HW-NAS-Bench:Hardware-Aware Neural Architecture Search Benchmark
• Consensus Control for Decentralized Deep Learning
• Lossy Image Compression with Normalizing Flows
• Model-based Asynchronous Hyperparameter and Neural Architecture Search
• GuidedMixup: An Efficient Mixup Strategy Guided by Saliency Maps
• STEERAGE: Synthesis of Neural Networks Using Architecture Search and Grow-and-Prune Methods
• Neural Ensemble Search for Uncertainty Estimation and Dataset Shift
• Co-Mixup: Saliency Guided Joint Mixup with Supermodular Diversity
• Using a thousand optimization tasks to learn hyperparameter search strategies
• Boosting Discriminative Visual Representation Learning with Scenario-Agnostic Mixup
• Stronger NAS with Weaker Predictors
• Rethinking the backbone architecture for tiny object detection
• Network Architecture Search for Domain Adaptation
• DVOLVER: Efficient Pareto-Optimal Neural Network Architecture Search
• Cyclic Differentiable Architecture Search
• ATOM: Robustifying Out-of-distribution Detection Using Outlier Mining
• Contrastive Learning Improves Model Robustness Under Label Noise
• Invertible DenseNets with Concatenated LipSwish
• Invertible Generative Modeling using Linear Rational Splines
• An Introduction to Neural Architecture Search for Convolutional Networks
• Depthwise Multiception Convolution for Reducing Network Parameters without Sacrificing Accuracy
• The Intrinsic Manifolds of Radiological Images and their Role in Deep Learning
• Local and Global Context-and-Object-part-Aware Superpixel-based Data Augmentation for Deep Visual Recognition
• Learning Rates as a Function of Batch Size: A Random Matrix Theory Approach to Neural Network Training
• Adaptive Consistency Regularization for Semi-Supervised Transfer Learning
• Instance Correction for Learning with Open-set Noisy Labels
• Stable, Fast and Accurate: Kernelized Attention with Relative Positional Encoding
• DrNAS: Dirichlet Neural Architecture Search
• Understanding and Accelerating Neural Architecture Search with Training-Free and Theory-Grounded Metrics
• Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN
• Representation Learning via Consistent Assignment of Views to Clusters
• Increasing Depth Leads to U-Shaped Test Risk in Over-parameterized Convolutional Networks
• Sample Selection Using Multi-Task Autoencoders in Federated Learning with Non-IID Data
• Energy-efficient Amortized Inference with Cascaded Deep Classifiers
• Trainless Model Performance Estimation for Neural Architecture Search
• FEAR: A Simple Lightweight Method to Rank Architectures
• Compressing gradients by exploiting temporal correlation in momentum-SGD
• Learnable Uncertainty under Laplace Approximations
• Being a Bit Frequentist Improves Bayesian Neural Networks
• BORE: Bayesian Optimization by Density-Ratio Estimation
• Learning Non-linear Wavelet Transformation via Normalizing Flow
• Neural Architecture Search with Random Labels
• An Infinite-Feature Extension for Bayesian ReLU Nets That Fixes Their Asymptotic Overconfidence
• The Unreasonable Effectiveness of Patches in Deep Convolutional Kernels Methods
• OSOA: One-Shot Online Adaptation of Deep Generative Models for Lossless Compression
• Confusable Learning for Large-class Few-Shot Classification
• Adversarial AutoMixup
• Factorized Neural Processes for Neural Processes: $K$-Shot Prediction of Neural Responses
• ExCon: Explanation-driven Supervised Contrastive Learning for Image Classification
• Human Annotations Improve GAN Performances
• GenURL: A General Framework for Unsupervised Representation Learning
• Sifting out the features by pruning: Are convolutional networks the winning lottery ticket of fully connected ones?
• Intra-Model Collaborative Learning of Neural Networks
• Midpoint Regularization: from High Uncertainty Training to Conservative Classification
• Lossless Image Compression Using a Multi-Scale Progressive Statistical Model
• Graphs for deep learning representations
• Distilling Double Descent
• Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator
• IB-DRR: Incremental Learning with Information-Back Discrete Representation Replay
• Adversarial Training with Stochastic Weight Average
• Training CNNs faster with Dynamic Input and Kernel Downsampling
• Signed Input Regularization
• Analyzing the Dependency of ConvNets on Spatial Information
• End-to-End Efficient Representation Learning via Cascading Combinatorial Optimization
• AmoebaContact and GDFold: a new pipeline for rapid prediction of protein structures