pith. sign in

arxiv: 2507.23315 · v2 · submitted 2025-07-31 · 💻 cs.CV · cs.AI· cs.LG

Analysis of Hyperparameter Optimization Effects on Lightweight Deep Models for Real-Time Image Classification

Vineet Kumar Rakesh , Soumya Mazumdar , Tapas Samanta , Hemendra Kumar Pandey , Amitabha Das This is my paper

Pith reviewed 2026-05-19 02:36 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords hyperparameter optimizationlightweight deep modelsreal-time image classificationImageNet subsetedge AI deploymentCNN transformersinference latency
0
0 comments X

The pith

Hyperparameter tuning improves accuracy of lightweight models by 1.5 to 3.5 percent while enabling high-speed inference.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines the effects of optimizing hyperparameters such as learning rate schedules, data augmentation, optimizers, and initialization on seven lightweight deep learning architectures for image classification. These models are trained on a balanced subset of 90,000 images from ImageNet-1K under standardized conditions. The results indicate that careful tuning enhances model convergence and boosts top-1 accuracy by 1.5 to 3.5 percent compared to baseline settings. Additionally, models like RepVGG-A2 and MobileNetV3-L demonstrate inference speeds suitable for real-time applications on edge devices, with latencies under 5 milliseconds and throughputs exceeding 9,800 frames per second on an NVIDIA L40s GPU.

Core claim

Under standardized training settings on a class-balanced 90,000-image subset of ImageNet-1K, controlled hyperparameter variation significantly alters convergence dynamics in lightweight CNN and transformer backbones. Tuning leads to a top-1 accuracy improvement of 1.5 to 3.5 percent over baselines. Select models such as RepVGG-A2 and MobileNetV3-L deliver latency under 5 milliseconds and over 9,800 frames per second, supporting deployment feasibility in edge artificial intelligence.

What carries the argument

The evaluation of hyperparameter effects including learning rate schedules, augmentation strategies, optimizers, and initialization on convergence and inference performance of lightweight models.

If this is right

  • Lightweight architectures can reach competitive accuracy levels through tuning rather than architectural changes.
  • High throughput models become viable for real-time edge deployment with proper hyperparameter selection.
  • Reproducible subset-based experiments provide guidance for balancing accuracy and speed in practical applications.
  • Insights into stability regions help in selecting models for resource-constrained environments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar tuning benefits might apply to other datasets if the class balance is maintained.
  • Combining these tuned models with quantization or pruning could further reduce latency for even stricter real-time constraints.
  • Future work could test these findings on full ImageNet or specialized datasets like medical imaging to check generalizability.

Load-bearing premise

The class-balanced 90,000-image subset of ImageNet-1K with the standardized training protocol is representative of broader convergence dynamics and deployment scenarios for lightweight models.

What would settle it

Re-training the models on the full ImageNet-1K dataset or a different large-scale dataset and observing no accuracy gains from the same hyperparameter tuning would indicate the subset results do not generalize.

Figures

Figures reproduced from arXiv: 2507.23315 by Amitabha Das, Hemendra Kumar Pandey, Soumya Mazumdar, Tapas Samanta, Vineet Kumar Rakesh.

Figure 1
Figure 1. Figure 1: Workflow of hyperparameter optimization for lightweight image classification models [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Accuracy vs Learning Rate for all evaluated models under consistent training settings. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Validation Accuracy vs Epochs for all models under a fixed training schedule of 300 epochs. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of Baseline and Optimized Validation Accuracy across models [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
read the original abstract

Lightweight convolutional and transformer-based networks are increasingly preferred for real-time image classification, especially on resource-constrained devices. This study evaluates the impact of hyperparameter optimization on the accuracy and deployment feasibility of seven modern lightweight architectures: ConvNeXt-T, EfficientNetV2-S, MobileNetV3-L, MobileViT v2 (S/XS), RepVGG-A2, and TinyViT-21M, trained on a class-balanced subset of 90,000 images from ImageNet-1K. Under standardized training settings, this paper investigates the influence of learning rate schedules, augmentation, optimizers, and initialization on model performance. Inference benchmarks are performed using an NVIDIA L40s GPU with batch sizes ranging from 1 to 512, capturing latency and throughput in real-time conditions. This work demonstrates that controlled hyperparameter variation significantly alters convergence dynamics in lightweight CNN and transformer backbones, providing insight into stability regions and deployment feasibility in edge artificial intelligence. Our results reveal that tuning alone leads to a top-1 accuracy improvement of 1.5 to 3.5 percent over baselines, and select models (e.g., RepVGG-A2, MobileNetV3-L) deliver latency under 5 milliseconds and over 9,800 frames per second, making them ideal for edge deployment. This work provides reproducible, subset-based insights into lightweight hyperparameter tuning and its role in balancing speed and accuracy. The code and logs may be seen at: https://vineetkumarrakesh.github.io/lcnn-opt

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript evaluates the effects of hyperparameter optimization (learning rate schedules, augmentation, optimizers, initialization) on seven lightweight models—ConvNeXt-T, EfficientNetV2-S, MobileNetV3-L, MobileViT v2 (S/XS), RepVGG-A2, and TinyViT-21M—trained on a class-balanced 90,000-image subset of ImageNet-1K. It reports that tuning produces 1.5–3.5% top-1 accuracy gains over baselines and that RepVGG-A2 and MobileNetV3-L achieve <5 ms latency and >9,800 FPS on an NVIDIA L40s GPU (batch sizes 1–512), concluding these models are ideal for edge deployment. Inference benchmarks and convergence analysis are presented with a link to code and logs.

Significance. If the accuracy improvements from hyperparameter tuning prove robust under statistical validation, the study could supply actionable guidance for practitioners tuning lightweight CNN and transformer backbones. The explicit provision of code and logs supports reproducibility, which strengthens the work’s utility. The deployment-feasibility conclusions, however, rest on hardware that does not match the claimed use case.

major comments (2)
  1. [Abstract] Abstract: the claim that RepVGG-A2 and MobileNetV3-L are 'ideal for edge deployment' because they deliver latency under 5 ms and over 9,800 FPS is not supported by the reported experiments, which benchmark exclusively on an NVIDIA L40s data-center GPU; no results on mobile SoCs, embedded GPUs, NPUs, or with INT8 quantization are provided, so the edge-deployment portion of the central claim lacks direct evidence.
  2. [Results on accuracy] Results on accuracy: the stated 1.5–3.5% top-1 accuracy improvement from tuning is presented without error bars, standard deviations from multiple random seeds, or statistical significance tests, leaving open the possibility that observed gains fall within run-to-run variation and weakening the evidence that hyperparameter optimization is the causal driver.
minor comments (2)
  1. [Abstract] The description of the 90,000-image subset would be clearer if it specified the exact class-balancing procedure and the number of classes retained from ImageNet-1K.
  2. [Inference benchmarks] Inference-benchmark details should explicitly define whether reported latency includes data loading or preprocessing and whether throughput is measured at steady state.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's detailed feedback on our manuscript. We address each of the major comments below and describe the revisions we intend to make to strengthen the paper.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that RepVGG-A2 and MobileNetV3-L are 'ideal for edge deployment' because they deliver latency under 5 ms and over 9,800 FPS is not supported by the reported experiments, which benchmark exclusively on an NVIDIA L40s data-center GPU; no results on mobile SoCs, embedded GPUs, NPUs, or with INT8 quantization are provided, so the edge-deployment portion of the central claim lacks direct evidence.

    Authors: We agree that the experiments were performed on an NVIDIA L40s GPU and do not include direct measurements on mobile or embedded devices. In the revised version, we will modify the abstract and conclusion to state that these models achieve low latency and high throughput on a high-end GPU, indicating potential for real-time applications, and we will qualify the edge deployment claim by noting that further evaluation on target hardware such as mobile SoCs would be beneficial. This addresses the lack of direct evidence while preserving the reported results. revision: yes

  2. Referee: [Results on accuracy] Results on accuracy: the stated 1.5–3.5% top-1 accuracy improvement from tuning is presented without error bars, standard deviations from multiple random seeds, or statistical significance tests, leaving open the possibility that observed gains fall within run-to-run variation and weakening the evidence that hyperparameter optimization is the causal driver.

    Authors: The referee correctly identifies that our current presentation lacks statistical measures of variability. To strengthen the evidence, we will rerun the experiments with multiple random seeds (at least three) for the key models and report mean accuracy with standard deviations. We will also include statistical significance tests (e.g., paired t-tests) to demonstrate that the improvements are significant. These additions will be incorporated into the results section of the revised manuscript. revision: yes

Circularity Check

0 steps flagged

Empirical benchmarking study with no circular derivation

full rationale

This is a pure empirical benchmarking paper that trains seven lightweight models on a fixed 90k-image balanced ImageNet subset, varies hyperparameters under a standardized protocol, and reports measured top-1 accuracy deltas plus inference latency/throughput on an NVIDIA L40s GPU. No equations, first-principles derivations, or predictive models are present; all results are direct experimental outcomes. There are no self-definitional loops, fitted inputs relabeled as predictions, or load-bearing self-citations that reduce the central claims to the paper's own inputs by construction. The analysis is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the representativeness of the ImageNet subset and the fairness of the standardized training protocol; no new free parameters or invented entities are introduced.

axioms (2)
  • domain assumption A class-balanced 90,000-image subset of ImageNet-1K is sufficient to evaluate convergence and deployment properties of lightweight models.
    Invoked when training and reporting accuracy and speed results.
  • domain assumption Standardized training settings produce comparable convergence dynamics across the seven architectures.
    Stated when comparing the effect of hyperparameter changes.

pith-pipeline@v0.9.0 · 5835 in / 1356 out tokens · 48920 ms · 2026-05-19T02:36:24.003962+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 5 internal anchors

  1. [1]

    TinyViT: Fast Pretraining Distillation for Small Vision Transformers,

    Kaiyu Wu, Jingdong Zhang, Hanbo Peng, Mengchen Liu, Bowen Xiao, Jianbo Fu, and Lu Yuan, “TinyViT: Fast Pretraining Distillation for Small Vision Transformers,” inEuropean Conference on Computer Vision (ECCV) , 2022, pp. 68–85

    work page 2022

  2. [2]

    SGDR: Stochastic Gradient Descent with Warm Restarts

    Ilya Loshchilov and Frank Hutter, “SGDR: Stochastic Gradient Descent with Warm Restarts,”arXiv preprint arXiv:1608.03983, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  3. [3]

    Searching for MobileNetV3,

    Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al., “Searching for MobileNetV3,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2019, pp. 1314–1324

    work page 2019

  4. [4]

    MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer

    Sachin Mehta, and Mohammad Rastegari, “MobileViT: Light-weight, General-purpose, and Mobile- friendly Vision Transformer,”arXiv preprint arXiv:2110.02178 , 2021

    work page internal anchor Pith review arXiv 2021

  5. [5]

    Separable Self-attention for Mobile Vision Transformers (MobileViTv2),

    Sachin Mehta, Nam Nguyen, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi, “Separable Self-attention for Mobile Vision Transformers (MobileViTv2),”Transactions on Machine Learning Research, 2023

    work page 2023

  6. [6]

    ConvNeXt: Revisiting ResNets at Scale,

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie, “ConvNeXt: Revisiting ResNets at Scale,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022, pp. 4817–4827

    work page 2022

  7. [7]

    ImageNet Large Scale Visual Recognition Challenge,

    Olga Russakovsky, Jia Deng, Hao Su, et al., “ImageNet Large Scale Visual Recognition Challenge,” International Journal of Computer Vision (IJCV) , vol. 115, no. 3, pp. 211–252, 2015. 10 A preprint - September 13, 2025

    work page 2015

  8. [8]

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,

    Mingxing Tan and Quoc V. Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” inProceedings of the 36th International Conference on Machine Learning (ICML) , 2019, pp. 6105–6114

    work page 2019

  9. [9]

    Le , title =

    Mingxing Tan and Quoc V. Le, “EfficientNetV2-S [9]: Smaller Models and Faster Training,”arXiv preprint arXiv:2104.00298, 2021

    work page arXiv 2021

  10. [10]

    RepVGG: Making VGG-style ConvNets Great Again,

    Xiangyu Ding, Xudong Zhang, Ningning Ma, Jianping Han, Guiguang Ding, and Jian Sun, “RepVGG: Making VGG-style ConvNets Great Again,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2021, pp. 13733–13742

    work page 2021

  11. [11]

    CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo, “CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2019, pp. 6023–6032

    work page 2019

  12. [12]

    D., Zoph, B., Mane, D., Vasudevan, V., & Le, Q

    Cubuk, E. D., Zoph, B., Mane, D., Vasudevan, V., & Le, Q. V. (2019). AutoAugment: Learning augmentation policies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2019), pp. 113–123

    work page 2019

  13. [13]

    RandAugment: Practical Automated Data Augmentation with a Reduced Search Space,

    Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le, “RandAugment: Practical Automated Data Augmentation with a Reduced Search Space,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020, pp. 18613–18624

    work page 2020

  14. [14]

    SGDR: Stochastic Gradient Descent with Warm Restarts,

    Ilya Loshchilov and Frank Hutter, “SGDR: Stochastic Gradient Descent with Warm Restarts,” in Proceedings of the 5th International Conference on Learning Representations (ICLR) , 2017

    work page 2017

  15. [15]

    mixup: Beyond Empirical Risk Minimization,

    Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz, “mixup: Beyond Empirical Risk Minimization,” inProceedings of the 6th International Conference on Learning Representations (ICLR), 2018

    work page 2018

  16. [16]

    SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size

    Forrest N. Iandola, Song Han, Matthew W. Moskewicz, Khalid Ashraf, William J. Dally, and Kurt Keutzer, “SqueezeNet: AlexNet-level accuracy with 50× fewer parameters and<0.5MB model size,” arXiv preprint arXiv:1602.07360 , 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  17. [17]

    MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam, “MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,”arXiv preprint arXiv:1704.04861 , 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  18. [18]

    Mo- bileNetV2: Inverted Residuals and Linear Bottlenecks,

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen, “Mo- bileNetV2: Inverted Residuals and Linear Bottlenecks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018, pp. 4510–4520

    work page 2018

  19. [19]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    Alexey Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” inProceedings of the 9th International Conference on Learning Representations (ICLR) , 2021

    work page 2021

  20. [20]

    Rethinking the Inception Architecture for Computer Vision,

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna, “Rethinking the Inception Architecture for Computer Vision,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2016, pp. 2818–2826

    work page 2016

  21. [21]

    Decoupled Weight Decay Regularization,

    Ilya Loshchilov and Frank Hutter, “Decoupled Weight Decay Regularization,” inProceedings of the 7th International Conference on Learning Representations (ICLR) , 2019

    work page 2019

  22. [22]

    Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

    Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, et al., “Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour,”arXiv preprint arXiv:1706.02677, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    Optimized Convo- lutional Neural Network by Firefly Algorithm for Magnetic Resonance Image Classification of Glioma Brain Tumor Grade,

    Natalija Bacanin, Tamara Bezdan, Kannan Venkatachalam, and Fadi Al–Turjman, “Optimized Convo- lutional Neural Network by Firefly Algorithm for Magnetic Resonance Image Classification of Glioma Brain Tumor Grade,”Journal of Real-Time Image Processing , vol. 18, no. 4, pp. 1085–1098, 2021

    work page 2021

  24. [24]

    Explaining Decisions of a Lightweight Deep Neural Network for Real-Time Coronary Artery Disease Classification in Magnetic Resonance Imaging,

    Tariq Iqbal, Ahsan Khalid, and Irfan Ullah, “Explaining Decisions of a Lightweight Deep Neural Network for Real-Time Coronary Artery Disease Classification in Magnetic Resonance Imaging,”Journal of Real-Time Image Processing, vol. 21, 2024. Author contributions VKR contributed to the conceptualization of the study, validation of results, and manuscript re...

    work page 2024