TinyNet-A: Efficient Image Classification for Mobile & Edge Hardware

TinyNet-A: Efficient Image Classification for Mobile & Edge Hardware

TL;DR

TinyNet-A is a compact 6.2-million-parameter image classification model built for mobile and edge hardware with limited compute and latency budgets.

Distributed as timm/tinynet_a.in1k, it uses 192 × 192 image inputs and can produce 1,280-dimensional features for downstream vision tasks.

On ImageNet-1k, TinyNet-A reports 76.8% Top-1 and 93.3% Top-5 accuracy, increasing to 77.7% Top-1 and 93.5% Top-5 with RandAugment.

What the Model Is

TinyNet-A, now listed on AIOZ AI, is a lightweight image classification model designed for resource-constrained environments.

It is trained on ImageNet-1k and distributed through timm as timm/tinynet_a.in1k. Beyond classification, TinyNet-A can also serve as a feature extractor or backbone in a larger vision system.

It is not an object detector by itself, but its visual features can support lightweight detection architectures such as SSDLite.

How TinyNet-A Is Designed

It is derived from EfficientNet-B0 using a downsizing method called the tiny formula.

The tiny formula searches for a smaller configuration under a fixed compute budget. It uses Gaussian process regression across Pareto-front models to optimize input resolution, network depth, and network width.

The resulting architecture uses 16 mobile inverted residual bottlenecks with a 192 × 192 input resolution.

Inputs and Outputs

TinyNet-A takes images prepared at 192 × 192 resolution.

Depending on the workflow, it can provide:

  • Image classification: Category predictions for ImageNet-1k tasks
  • Feature extraction: A 1,280-dimensional visual representation
  • Detection backbone: Features for a separate detector such as SSDLite

Core Capabilities

  • Compact image classification
  • Mobile and edge inference
  • On-device image categorization
  • 1,280-dimensional feature extraction
  • Backbone support for lightweight detection
  • Efficient inference under constrained compute

Efficiency and ImageNet Performance

  • TinyNet-A compute: 339M FLOPs
  • EfficientNet-B0 compute: 387M FLOPs
  • Compute reduction: 12.4%
  • TinyNet-A latency on Huawei P40: 81.30 ms
  • EfficientNet-B0 latency on Huawei P40: 99.85 ms
  • TinyNet-A accuracy: 76.8% Top-1 / 93.3% Top-5
  • With RandAugment: 77.7% Top-1 / 93.5% Top-5

These results depend on the reported benchmark setup and hardware.

Key Technical Details

  • Model: TinyNet-A
  • Distribution: timm/tinynet_a.in1k
  • Task: Image classification
  • Parameters: 6.2M
  • Training dataset: ImageNet-1k
  • Input resolution: 192 × 192
  • Architecture origin: EfficientNet-B0
  • Scaling approach: Tiny formula
  • Optimization dimensions: Resolution, depth, width
  • Mobile inverted residual bottlenecks: 16
  • Feature dimension: 1,280
  • Reported compute: 339M FLOPs
  • License: Apache 2.0

Explore It on AIOZ AI

TinyNet-A combines a compact parameter count, reduced compute, and competitive ImageNet-1k performance for mobile and edge vision workflows.

Explore TinyNet-A on AIOZ AI and evaluate how it fits resource-constrained computer-vision applications.

FAQ

Q1: What is TinyNet-A?

It is a 6.2-million-parameter image classification model derived from EfficientNet-B0 and optimized for mobile and edge hardware.

Q2: Can TinyNet-A perform object detection by itself?

No. It can serve as a visual backbone for a detector such as SSDLite, while the separate detection head produces object classes and bounding boxes.

Q3: What can TinyNet-A output besides classification predictions?

It can produce a 1,280-dimensional feature representation for downstream tasks.