| --- |
| license: mit |
| datasets: |
| - zh-plus/tiny-imagenet |
| pipeline_tag: image-classification |
| library_name: transformers |
| tags: |
| - Squeeze |
| - New |
| - Image |
| - Clf |
| metrics: |
| - accuracy |
| --- |
| |
| # SqueezeNet-SwiGLU (Modernized SqueezeNet Architecture) |
|
|
| **SqueezeNet-SwiGLU** is a modernized, ultra-lightweight Convolutional Neural Network (CNN) architecture based on the original SqueezeNet v1.1 design. It incorporates state-of-the-art deep learning architectural enhancements, including **SwiGLU gated activations**, **FP32 Layer Normalization**, and **residual block scaling**, delivering superior feature representation while maintaining a minimal parameter footprint. |
|
|
| --- |
|
|
| ## Key Architectural Improvements (vs. Original SqueezeNet) |
|
|
| Compared to the legacy SqueezeNet v1.1 (Iandola et al., 2016), this model introduces several architectural modernizations: |
|
|
| | Feature | Legacy SqueezeNet v1.1 | **SqueezeNet-SwiGLU (This Model)** | |
| | :--- | :--- | :--- | |
| | **Activation Function** | Standard ReLU | **Residual-Scaled SwiGLU Gated Activation** | |
| | **Normalization** | Batch Normalization | **FP32 Layer Normalization (`GroupNorm(1, C)`)** | |
| | **Batch Size Dependency** | High (sensitive to batch stats & EMA lag) | **Zero (Inference identical across any batch size)** | |
| | **Gradient Flow** | Standard Fire Connections | **Residual Fire Block Skip Connections & Scaling** | |
| | **Activation Variance** | Prone to un-bounded drift | **Strictly bounded via LayerNorm & FP32 Precision** | |
|
|
| --- |
|
|
| ## Benchmark & Evaluation |
|
|
| - **Evaluation Dataset**: ImageNet 200-Class Test Split (Tiny-ImageNet Categories) |
| - **Input Resolution**: 64 × 64 pixels (native) / 224 × 224 (interpolated) |
| - **Top-1 Accuracy**: 50.51% |
| - **Top-5 Accuracy**: 75.06% |
|
|
| --- |
|
|
| ## Target Usecases & Applications |
|
|
| Due to its ultra-compact size and high throughput, **SqueezeNet-SwiGLU** is optimized for resource-constrained deployment environments: |
|
|
| 1. **Edge & IoT Intelligence**: Microcontrollers, Raspberry Pi, NVIDIA Jetson, and embedded vision hardware. |
| 2. **Mobile AI Applications**: On-device real-time visual classification (iOS CoreML / Android ONNX). |
| 3. **High-FPS Video Analytics**: Lightweight feature backbone for real-time surveillance, robotics, and drone navigation. |
| 4. **Microservice Backends**: Serving high-throughput image classification with minimal memory overhead per GPU/CPU node. |
|
|
| --- |
|
|
| ## How to Use |
|
|
| ### Fast Inference with Hugging Face `pipeline` |
|
|
| ```python |
| from transformers import pipeline |
| |
| # Initialize the classification pipeline (requires trust_remote_code=True) |
| classifier = pipeline( |
| "image-classification", |
| model="kd13/Modern-SqueezeNet", |
| 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}") |
