| Property | Value |
|---|---|
| Parameter Count | 71.5M |
| License | Apache-2.0 |
| Framework | PyTorch (timm) |
| Training Data | ImageNet-1k |
| Paper | High-Performance Large-Scale Image Recognition Without Normalization |
What is dm_nfnet_f0.dm_in1k?
dm_nfnet_f0.dm_in1k is a Normalization Free Network (NFNet) designed for image classification tasks. This innovative model architecture eliminates the need for traditional normalization layers while maintaining high performance. Instead of using Batch Normalization, it employs Scaled Weight Standardization and strategically placed scalar gains in the residual path.
Implementation Details
The model features a sophisticated architecture with 71.5M parameters and requires 7.2 GMACs for inference. It operates on image sizes of 192x192 during training and 256x256 during testing, utilizing modified ResNet-like structures without normalization layers. The model uses pre-activation designs and implements specific signal propagation techniques to achieve optimal performance.
- Uses Scaled Weight Standardization instead of traditional normalization
- Implements strategic scalar gains in residual paths
- Features pre-activation ResNet-like architecture
- Supports both classification and feature extraction modes
Core Capabilities
- Image Classification with 1000 classes (ImageNet)
- Feature Map Extraction with multiple resolution levels
- Image Embedding Generation
- Support for both inference and feature extraction workflows
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its ability to achieve high performance without using normalization layers, which traditionally have been considered essential for deep neural networks. It introduces novel techniques like Scaled Weight Standardization and carefully placed scalar gains to maintain stability and performance.
Q: What are the recommended use cases?
The model is best suited for large-scale image classification tasks, feature extraction for downstream tasks, and generating image embeddings. It's particularly effective when working with high-resolution images and when normalization-free architectures are preferred for computational efficiency.