cifar10-airbench

94% on CIFAR-10 in 2.67 seconds 💨 96% in 27 seconds

Stars: 147

Visit

CIFAR-10 Airbench is a project offering fast and stable training baselines for CIFAR-10 dataset, facilitating machine learning research. It provides easily runnable PyTorch scripts for training neural networks with high accuracy levels. The methods used in this project aim to accelerate research on fundamental properties of deep learning. The project includes GPU-accelerated dataloader for custom experiments and trainings, and can be used for data selection and active learning experiments. The training methods provided are faster than standard ResNet training, offering improved performance for research projects.

README:

CIFAR-10 Airbench 💨

A collection of fast and self-contained training scripts for CIFAR-10.

Script	Mean accuracy	Time	PFLOPs
`airbench94_compiled.py`	94.01%	3.09s	0.36
`airbench94.py`	94.01%	3.83s	0.36
`airbench95.py`	95.01%	10.4s	1.4
`airbench96.py`	96.03%	34.7s	4.9
`airbench94_muon.py`	94.05%	2.67s	0.29
`airbench96_faster.py`	96.00%	27.3s	3.1

For a comparison, the standard training used in most studies on CIFAR-10 is much slower:

Baseline	Mean accuracy	Time	PFLOPs
Standard ResNet-18 training	96%	7min	32.3

All timings are on a single NVIDIA A100 GPU.

Note: airbench96 has been improved since the paper from 46s to 35s. In addition, airbench96_faster is an improved (but more complicated) method which uses data filtering by a small proxy model. And airbench94_muon is an improved method using a variant of the Muon optimizer.

The set of methods used to obtain these training speeds are described in the paper.

How to run

To train a neural network with 94% accuracy, run either

git clone https://github.com/KellerJordan/cifar10-airbench.git
cd airbench && python airbench94.py

pip install airbench
python -c "import airbench; airbench.warmup94(); airbench.train94()"

Note: airbench94_compiled.py and airbench94.py are equivalent (i.e., yield the same distribution of trained networks), and differ only in that the first uses torch.compile to improve GPU utilization. The former is intended for experiments where many networks are trained at once in order to amortize the one-time compilation cost.

Motivation

CIFAR-10 is one of the most widely used datasets in machine learning, facilitating thousands of research projects per year. This repo provides fast and stable training baselines for CIFAR-10 in order to help accelerate this research. The trainings are provided as easily runnable dependency-free PyTorch scripts, and can replace classic baselines like training ResNet-20 or ResNet-18.

Using the GPU-accelerated dataloader independently

For writing custom CIFAR-10 experiments or trainings, you may find it useful to use the GPU-accelerated dataloader independently.

import airbench
train_loader = airbench.CifarLoader('/tmp/cifar10', train=True, aug=dict(flip=True, translate=4, cutout=16), batch_size=500)
test_loader = airbench.CifarLoader('/tmp/cifar10', train=False, batch_size=1000)

for epoch in range(200):
    for inputs, labels in train_loader:
        # outputs = model(inputs)
        # loss = F.cross_entropy(outputs, labels)
        ...

If you wish to modify the data in the loader, it can be done like so:

import airbench
train_loader = airbench.CifarLoader('/tmp/cifar10', train=True, aug=dict(flip=True, translate=4, cutout=16), batch_size=500)
mask = (train_loader.labels < 6) # (this is just an example, the mask can be anything)
train_loader.images = train_loader.images[mask]
train_loader.labels = train_loader.labels[mask]
print(len(train_loader)) # The loader now contains 30,000 images and has batch size 500, so this prints 60.

Example data-selection experiment

Airbench can be used as a platform for experiments in data selection and active learning. The following is an example experiment which demonstrates the classic result that low-confidence examples provide more training signal than random examples. It runs in <20 seconds on an A100.

import torch
from airbench import train94, infer, evaluate, CifarLoader

net = train94(label_smoothing=0) # train this network without label smoothing to get a better confidence signal

loader = CifarLoader('cifar10', train=True, batch_size=1000)
logits = infer(net, loader)
conf = logits.log_softmax(1).amax(1) # confidence

train_loader = CifarLoader('cifar10', train=True, batch_size=1024, aug=dict(flip=True, translate=2))
mask = (torch.rand(len(train_loader.labels)) < 0.6)
print('Training on %d images selected randomly' % mask.sum())
train_loader.images = train_loader.images[mask]
train_loader.labels = train_loader.labels[mask]
train94(train_loader, epochs=16) # yields around 93% accuracy

train_loader = CifarLoader('cifar10', train=True, batch_size=1024, aug=dict(flip=True, translate=2))
mask = (conf < conf.float().quantile(0.6))
print('Training on %d images selected based on minimum confidence' % mask.sum())
train_loader.images = train_loader.images[mask]
train_loader.labels = train_loader.labels[mask]
train94(train_loader, epochs=16) # yields around 94% accuracy => low-confidence sampling is better than random.

Prior work

This project builds on the excellent previous record https://github.com/tysam-code/hlb-CIFAR10 (6.3 A100-seconds to 94%).

Which itself builds on the amazing series https://myrtle.ai/learn/how-to-train-your-resnet/ (26 V100-seconds to 94%, which is >=8 A100-seconds)

For Tasks:

Click tags to check more tools for each tasks

train neural networks conduct data experiments accelerate research perform active learning improve deep learning models

For Jobs:

machine learning engineer data scientist research scientist ai engineer deep learning researcher

Alternative AI tools for cifar10-airbench

Similar Open Source Tools

cifar10-airbench

github

: 147

CodeGeeX4

CodeGeeX4-ALL-9B is an open-source multilingual code generation model based on GLM-4-9B, offering enhanced code generation capabilities. It supports functions like code completion, code interpreter, web search, function call, and repository-level code Q&A. The model has competitive performance on benchmarks like BigCodeBench and NaturalCodeBench, outperforming larger models in terms of speed and performance.

github

: 1.0k

cladder

CLadder is a repository containing the CLadder dataset for evaluating causal reasoning in language models. The dataset consists of yes/no questions in natural language that require statistical and causal inference to answer. It includes fields such as question_id, given_info, question, answer, reasoning, and metadata like query_type and rung. The dataset also provides prompts for evaluating language models and example questions with associated reasoning steps. Additionally, it offers dataset statistics, data variants, and code setup instructions for using the repository.

github

: 82

GPTQModel

GPTQModel is an easy-to-use LLM quantization and inference toolkit based on the GPTQ algorithm. It provides support for weight-only quantization and offers features such as dynamic per layer/module flexible quantization, sharding support, and auto-heal quantization errors. The toolkit aims to ensure inference compatibility with HF Transformers, vLLM, and SGLang. It offers various model supports, faster quant inference, better quality quants, and security features like hash check of model weights. GPTQModel also focuses on faster quantization, improved quant quality as measured by PPL, and backports bug fixes from AutoGPTQ.

github

: 104

litdata

LitData is a tool designed for blazingly fast, distributed streaming of training data from any cloud storage. It allows users to transform and optimize data in cloud storage environments efficiently and intuitively, supporting various data types like images, text, video, audio, geo-spatial, and multimodal data. LitData integrates smoothly with frameworks such as LitGPT and PyTorch, enabling seamless streaming of data to multiple machines. Key features include multi-GPU/multi-node support, easy data mixing, pause & resume functionality, support for profiling, memory footprint reduction, cache size configuration, and on-prem optimizations. The tool also provides benchmarks for measuring streaming speed and conversion efficiency, along with runnable templates for different data types. LitData enables infinite cloud data processing by utilizing the Lightning.ai platform to scale data processing with optimized machines.

github

: 340

rl

TorchRL is an open-source Reinforcement Learning (RL) library for PyTorch. It provides pytorch and **python-first** , low and high level abstractions for RL that are intended to be **efficient** , **modular** , **documented** and properly **tested**. The code is aimed at supporting research in RL. Most of it is written in python in a highly modular way, such that researchers can easily swap components, transform them or write new ones with little effort.

github

: 2.3k

ABQ-LLM

ABQ-LLM is a novel arbitrary bit quantization scheme that achieves excellent performance under various quantization settings while enabling efficient arbitrary bit computation at the inference level. The algorithm supports precise weight-only quantization and weight-activation quantization. It provides pre-trained model weights and a set of out-of-the-box quantization operators for arbitrary bit model inference in modern architectures.

github

: 99

Qwen

Qwen is a series of large language models developed by Alibaba DAMO Academy. It outperforms the baseline models of similar model sizes on a series of benchmark datasets, e.g., MMLU, C-Eval, GSM8K, MATH, HumanEval, MBPP, BBH, etc., which evaluate the models’ capabilities on natural language understanding, mathematic problem solving, coding, etc. Qwen models outperform the baseline models of similar model sizes on a series of benchmark datasets, e.g., MMLU, C-Eval, GSM8K, MATH, HumanEval, MBPP, BBH, etc., which evaluate the models’ capabilities on natural language understanding, mathematic problem solving, coding, etc. Qwen-72B achieves better performance than LLaMA2-70B on all tasks and outperforms GPT-3.5 on 7 out of 10 tasks.

github

: 13.5k

beyondllm

Beyond LLM offers an all-in-one toolkit for experimentation, evaluation, and deployment of Retrieval-Augmented Generation (RAG) systems. It simplifies the process with automated integration, customizable evaluation metrics, and support for various Large Language Models (LLMs) tailored to specific needs. The aim is to reduce LLM hallucination risks and enhance reliability.

github

: 254

FlexFlow

FlexFlow Serve is an open-source compiler and distributed system for **low latency**, **high performance** LLM serving. FlexFlow Serve outperforms existing systems by 1.3-2.0x for single-node, multi-GPU inference and by 1.4-2.4x for multi-node, multi-GPU inference.

github

: 1.7k

Consistency_LLM

Consistency Large Language Models (CLLMs) is a family of efficient parallel decoders that reduce inference latency by efficiently decoding multiple tokens in parallel. The models are trained to perform efficient Jacobi decoding, mapping any randomly initialized token sequence to the same result as auto-regressive decoding in as few steps as possible. CLLMs have shown significant improvements in generation speed on various tasks, achieving up to 3.4 times faster generation. The tool provides a seamless integration with other techniques for efficient Large Language Model (LLM) inference, without the need for draft models or architectural modifications.

github

: 293

pywhy-llm

PyWhy-LLM is an innovative library that integrates Large Language Models (LLMs) into the causal analysis process, empowering users with knowledge previously only available through domain experts. It seamlessly augments existing causal inference processes by suggesting potential confounders, relationships between variables, backdoor sets, front door sets, IV sets, estimands, critiques of DAGs, latent confounders, and negative controls. By leveraging LLMs and formalizing human-LLM collaboration, PyWhy-LLM aims to enhance causal analysis accessibility and insight.

github

: 60

multipack_sampler

The Multipack sampler is a tool designed for padding-free distributed training of large language models. It optimizes batch processing efficiency using an approximate solution to the identical machine scheduling problem. The V2 update further enhances the packing algorithm complexity, achieving better throughput for a large number of nodes. It includes two variants for models with different attention types, aiming to balance sequence lengths and optimize packing efficiency. Users can refer to the provided benchmark for evaluating efficiency, utilization, and L^2 lag. The tool is compatible with PyTorch DataLoader and is released under the MIT license.

github

: 162

zeta

Zeta is a tool designed to build state-of-the-art AI models faster by providing modular, high-performance, and scalable building blocks. It addresses the common issues faced while working with neural nets, such as chaotic codebases, lack of modularity, and low performance modules. Zeta emphasizes usability, modularity, and performance, and is currently used in hundreds of models across various GitHub repositories. It enables users to prototype, train, optimize, and deploy the latest SOTA neural nets into production. The tool offers various modules like FlashAttention, SwiGLUStacked, RelativePositionBias, FeedForward, BitLinear, PalmE, Unet, VisionEmbeddings, niva, FusedDenseGELUDense, FusedDropoutLayerNorm, MambaBlock, Film, hyper_optimize, DPO, and ZetaCloud for different tasks in AI model development.

github

: 365

SwanLab

SwanLab is an open-source, lightweight AI experiment tracking tool that provides a platform for tracking, comparing, and collaborating on experiments, aiming to accelerate the research and development efficiency of AI teams by 100 times. It offers a friendly API and a beautiful interface, combining hyperparameter tracking, metric recording, online collaboration, experiment link sharing, real-time message notifications, and more. With SwanLab, researchers can document their training experiences, seamlessly communicate and collaborate with collaborators, and machine learning engineers can develop models for production faster.

github

: 433

zipnn

ZipNN is a lossless and near-lossless compression library optimized for numbers/tensors in the Foundation Models environment. It automatically prepares data for compression based on its type, allowing users to focus on core tasks without worrying about compression complexities. The library delivers effective compression techniques for different data types and structures, achieving high compression ratios and rates. ZipNN supports various compression methods like ZSTD, lz4, and snappy, and provides ready-made scripts for file compression/decompression. Users can also manually import the package to compress and decompress data. The library offers advanced configuration options for customization and validation tests for different input and compression types.

github

: 151

For similar tasks

cifar10-airbench

github

: 147

nntrainer

NNtrainer is a software framework for training neural network models on devices with limited resources. It enables on-device fine-tuning of neural networks using user data for personalization. NNtrainer supports various machine learning algorithms and provides examples for tasks such as few-shot learning, ResNet, VGG, and product rating. It is optimized for embedded devices and utilizes CBLAS and CUBLAS for accelerated calculations. NNtrainer is open source and released under the Apache License version 2.0.

github

: 135

uvadlc_notebooks

The UvA Deep Learning Tutorials repository contains a series of Jupyter notebooks designed to help understand theoretical concepts from lectures by providing corresponding implementations. The notebooks cover topics such as optimization techniques, transformers, graph neural networks, and more. They aim to teach details of the PyTorch framework, including PyTorch Lightning, with alternative translations to JAX+Flax. The tutorials are integrated as official tutorials of PyTorch Lightning and are relevant for graded assignments and exams.

github

: 2.5k

Deej-AI

Deej-A.I. is an advanced machine learning project that aims to revolutionize music recommendation systems by using artificial intelligence to analyze and recommend songs based on their content and characteristics. The project involves scraping playlists from Spotify, creating embeddings of songs, training neural networks to analyze spectrograms, and generating recommendations based on similarities in music features. Deej-A.I. offers a unique approach to music curation, focusing on the 'what' rather than the 'how' of DJing, and providing users with personalized and creative music suggestions.

github

: 312

awesome-ai

Awesome AI is a curated list of artificial intelligence resources including courses, tools, apps, and open-source projects. It covers a wide range of topics such as machine learning, deep learning, natural language processing, robotics, conversational interfaces, data science, and more. The repository serves as a comprehensive guide for individuals interested in exploring the field of artificial intelligence and its applications across various domains.

github

: 453

netsaur

Netsaur is a powerful machine learning library for Deno, offering a lightweight and easy-to-use neural network solution. It is blazingly fast and efficient, providing a simple API for creating and training neural networks. Netsaur can run on both CPU and GPU, making it suitable for serverless environments. With Netsaur, users can quickly build and deploy machine learning models for various applications with minimal dependencies. This library is perfect for both beginners and experienced machine learning practitioners.

github

: 211

For similar jobs

sweep

Sweep is an AI junior developer that turns bugs and feature requests into code changes. It automatically handles developer experience improvements like adding type hints and improving test coverage.

github

: 7.1k

teams-ai

The Teams AI Library is a software development kit (SDK) that helps developers create bots that can interact with Teams and Microsoft 365 applications. It is built on top of the Bot Framework SDK and simplifies the process of developing bots that interact with Teams' artificial intelligence capabilities. The SDK is available for JavaScript/TypeScript, .NET, and Python.

github

: 390

ai-guide

This guide is dedicated to Large Language Models (LLMs) that you can run on your home computer. It assumes your PC is a lower-end, non-gaming setup.

github

: 159

classifai

Supercharge WordPress Content Workflows and Engagement with Artificial Intelligence. Tap into leading cloud-based services like OpenAI, Microsoft Azure AI, Google Gemini and IBM Watson to augment your WordPress-powered websites. Publish content faster while improving SEO performance and increasing audience engagement. ClassifAI integrates Artificial Intelligence and Machine Learning technologies to lighten your workload and eliminate tedious tasks, giving you more time to create original content that matters.

github

: 572

chatbot-ui

Chatbot UI is an open-source AI chat app that allows users to create and deploy their own AI chatbots. It is easy to use and can be customized to fit any need. Chatbot UI is perfect for businesses, developers, and anyone who wants to create a chatbot.

github

: 27.7k

BricksLLM

BricksLLM is a cloud native AI gateway written in Go. Currently, it provides native support for OpenAI, Anthropic, Azure OpenAI and vLLM. BricksLLM aims to provide enterprise level infrastructure that can power any LLM production use cases. Here are some use cases for BricksLLM: * Set LLM usage limits for users on different pricing tiers * Track LLM usage on a per user and per organization basis * Block or redact requests containing PIIs * Improve LLM reliability with failovers, retries and caching * Distribute API keys with rate limits and cost limits for internal development/production use cases * Distribute API keys with rate limits and cost limits for students

github

: 897

uAgents

uAgents is a Python library developed by Fetch.ai that allows for the creation of autonomous AI agents. These agents can perform various tasks on a schedule or take action on various events. uAgents are easy to create and manage, and they are connected to a fast-growing network of other uAgents. They are also secure, with cryptographically secured messages and wallets.

github

: 951

griptape

Griptape is a modular Python framework for building AI-powered applications that securely connect to your enterprise data and APIs. It offers developers the ability to maintain control and flexibility at every step. Griptape's core components include Structures (Agents, Pipelines, and Workflows), Tasks, Tools, Memory (Conversation Memory, Task Memory, and Meta Memory), Drivers (Prompt and Embedding Drivers, Vector Store Drivers, Image Generation Drivers, Image Query Drivers, SQL Drivers, Web Scraper Drivers, and Conversation Memory Drivers), Engines (Query Engines, Extraction Engines, Summary Engines, Image Generation Engines, and Image Query Engines), and additional components (Rulesets, Loaders, Artifacts, Chunkers, and Tokenizers). Griptape enables developers to create AI-powered applications with ease and efficiency.

github

: 2.0k