NFNet-L0 (RA2)
| Property | Value |
|---|---|
| Parameter Count | 35.1M |
| License | Apache 2.0 |
| Training Data | ImageNet-1K |
| GMACs | 4.4 |
| Image Size | Train: 224x224, Test: 288x288 |
What is nfnet_l0.ra2_in1k?
NFNet-L0 is a lightweight variant of the Normalization-Free Network architecture, developed as part of the timm library. This model represents a significant innovation in deep learning by eliminating traditional normalization layers while maintaining high performance through scaled weight standardization.
Implementation Details
The model employs a unique architecture that replaces conventional batch normalization with carefully placed scalar gains in the residual path. It features reduced SE and bottleneck ratios (0.25 instead of 0.5) and uses SiLU activations instead of GELU, making it more efficient while maintaining performance.
- Lightweight design with 35.1M parameters
- Scaled Weight Standardization implementation
- Modified residual pathways with strategic scalar gains
- Optimized for both training (224x224) and testing (288x288) resolutions
Core Capabilities
- Image Classification on ImageNet-1K
- Feature Map Extraction
- Image Embedding Generation
- Flexible backbone for transfer learning
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its normalization-free architecture, which achieves high performance without traditional batch normalization layers, using scaled weight standardization instead. It's specifically optimized for efficiency while maintaining robust performance.
Q: What are the recommended use cases?
The model is well-suited for image classification tasks, feature extraction, and as a backbone for transfer learning applications. It's particularly effective when deployment efficiency is a priority while maintaining high accuracy.
