metadata
license: mit
datasets:
- zh-plus/tiny-imagenet
metrics:
- accuracy
pipeline_tag: image-classification
library_name: transformers
tags:
- Mobile
- edge
- image
- clf
Modern MobileNetV1 (Modernized MobileNet Architecture)
Modern MobileNetV1 is an enhanced, highly optimized variant of the classic MobileNetV1 architecture. It incorporates modern deep learning design choices—including SiLU activations, FP32 Layer Normalization, and learnable residual scaling—delivering stabilized training and high inference accuracy while keeping memory footprint and computational complexity low.
Key Architectural Improvements (vs. Original MobileNetV1)
Compared to the classic MobileNetV1 (Howard et al., 2017), this modernized implementation introduces several key architectural upgrades:
| Feature | Legacy MobileNetV1 | Modern MobileNetV1 (This Model) |
|---|---|---|
| Activation Function | Standard ReLU | SiLU (Swish) |
| Normalization | Batch Normalization | FP32 Layer Normalization (GroupNorm(1, C)) |
| Residual Connections | None (pure feed-forward) | Learnable Residual Block Scaling (identity + scale * out) |
| Batch Size Dependency | High (sensitive to batch statistics) | Zero (Inference identical across any batch size) |
| Precision Stability | Standard FP32 / FP16 | FP32-Capped Normalization (Prevents Underflow/Overflow) |
Benchmark & Evaluation
- Evaluation Dataset: Tiny-ImageNet (200-Class Test Split)
- Input Resolution: 64 × 64 pixels (native)
- Top-1 Accuracy: 44.38%
- Top-5 Accuracy: 67.26%
Target Use Cases & Applications
Due to its parameter efficiency and depthwise separable convolution structure, Modern MobileNetV1 is optimized for edge deployment:
- Edge & Embedded AI: Deployment on Raspberry Pi, NVIDIA Jetson, microcontrollers, and IoT vision devices.
- Mobile Vision Applications: Real-time on-device classification (Android ONNX / iOS CoreML).
- High-Throughput Microservices: Lightweight backbone for low-latency web services and microservices.
- Robotics & Drones: Compact feature extractor for fast object recognition and navigational awareness.
How to Use
Fast Inference with Hugging Face pipeline
from transformers import pipeline
# Initialize the classification pipeline (requires trust_remote_code=True for custom code)
classifier = pipeline(
"image-classification",
model="kd13/Modern-MobileNet",
trust_remote_code=True
)
# Run prediction on an image URL or local PIL Image
results = classifier("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/parrots.png")
for pred in results:
print(f"Label: {pred['label']} | Score: {pred['score']:.4f}")