| Property | Value |
|---|---|
| Parameter Count | 10.8M |
| Model Type | Image Classification |
| Architecture | HaloNet with ECA |
| License | Apache-2.0 |
| Paper | Scaling Local Self-Attention |
What is eca_halonext26ts.c1_in1k?
eca_halonext26ts.c1_in1k is a sophisticated image classification model that combines HaloNet architecture with Efficient Channel Attention (ECA), built on ResNeXt principles. Developed by Ross Wightman in the timm framework, this model achieves efficient performance with just 10.8M parameters while processing 256x256 images.
Implementation Details
The model is implemented using timm's flexible BYOBNet (Bring-Your-Own-Blocks Network) framework, incorporating advanced training techniques from the "ResNet Strikes Back" methodology. It uses SGD with Nesterov momentum and adaptive gradient clipping, along with a cosine learning rate schedule with warmup.
- GMACs: 2.4
- Activations: 11.5M
- Input Resolution: 256x256
- Trained on ImageNet-1k dataset
Core Capabilities
- Image classification with state-of-the-art accuracy
- Feature extraction for downstream tasks
- Efficient channel attention mechanism
- Flexible architecture with customizable block layouts
- Support for gradient checkpointing and layer-wise LR decay
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its combination of HaloNet's local self-attention mechanism with efficient channel attention, optimized for reduced frequency of self-attention blocks while maintaining performance. It's particularly notable for its parameter efficiency while handling high-resolution images.
Q: What are the recommended use cases?
The model is ideal for image classification tasks, feature extraction, and as a backbone for computer vision applications. It's particularly suitable for scenarios requiring a good balance between computational efficiency and accuracy.