[Paper Review] Evaluating ResNeXt Model Architecture for Image Classification

Saifuddin Hitawala|arXiv (Cornell University)|May 9, 2018

COVID-19 diagnosis using AI14 references16 citations

TL;DR

This paper evaluates the ResNeXt architecture on CIFAR-10 subsets by modifying key hyperparameters—cardinality, depth, and base-width—using PyTorch. Results show that reducing depth from 29 to 20 or base-width from 64 to 32 yields comparable accuracy with significantly reduced training time, indicating viable trade-offs for efficient model training without major performance loss.

ABSTRACT

In recent years, deep learning methods have been successfully applied to image classification tasks. Many such deep neural networks exist today that can easily differentiate cats from dogs. One such model is the ResNeXt model that uses a homogeneous, multi-branch architecture for image classification. This paper aims at implementing and evaluating the ResNeXt model architecture on subsets of the CIFAR-10 dataset. It also tweaks the original ResNeXt hyper-parameters such as cardinality, depth and base-width and compares the performance of the modified model with the original. Analysis of the experiments performed in this paper show that a slight decrease in depth or base-width does not affect the performance of the model much leading to comparable results.

Motivation & Objective

• To evaluate the performance of the ResNeXt model architecture on subsets of the CIFAR-10 dataset.
• To investigate the impact of hyperparameter tuning—specifically cardinality, depth, and base-width—on model accuracy and training efficiency.
• To determine whether reduced model configurations can achieve comparable performance to the original ResNeXt with less computational cost.
• To analyze training stability and convergence behavior under modified hyperparameters using limited computational resources.

Proposed method

• Implemented the original ResNeXt model using PyTorch for training on CIFAR-10 subsets.
• Systematically varied one hyperparameter at a time—cardinality, depth, and base-width—while keeping all other settings constant.
• Trained models for 300 epochs using stochastic gradient descent with standard optimization settings.
• Monitored training and validation accuracy and loss across all configurations to assess performance and convergence.
• Used SHARCNET clusters for training, with checkpointing to manage job time limits of up to 3 hours.
• Compared final test accuracy, training time per epoch, and variance in test error across different configurations.

Experimental results

Research questions

• RQ1 How does reducing the depth of the ResNeXt model from 29 to 20 affect classification accuracy on CIFAR-10 subsets?
• RQ2 What is the impact of decreasing the base-width from 64 to 32 on model performance and training efficiency?
• RQ3 How does varying cardinality affect the model’s accuracy and convergence behavior on small-scale CIFAR-10 subsets?
• RQ4 Can simplified ResNeXt configurations achieve comparable accuracy to the original model while reducing training time?

Key findings

• Reducing the model depth from 29 to 20 resulted in a slight decrease in final test accuracy, but the variance in test error was significantly lower, indicating improved training stability.
• The training time for the depth-20 model was approximately 70 minutes, a reduction of about 30 minutes compared to the depth-29 model, while maintaining high accuracy.
• Decreasing the base-width from 64 to 32 reduced the number of parameters from 32.4 million to 22.8 million and cut training time to around 75 minutes, with comparable test accuracy to the base-width 64 model.
• The model with base-width 32 showed nearly identical convergence rates and test accuracy variance compared to the base-width 64 model, indicating robustness to width reduction.
• The best test accuracy achieved on the CIFAR-10 subset was 83%, closely matching the 83.7% reported in the original ResNeXt paper for CIFAR-100, suggesting that dataset size and class distribution heavily influence performance.
• Inconsistent training behavior in some runs (e.g., Cifar-2 with cardinality 4 and Cifar-10 with cardinality 16) was attributed to improper checkpoint resumption due to limited job time on SHARCNET.