| Property | Value |
|---|---|
| Parameter Count | 10.6M |
| Model Type | Image Classification |
| License | Apache-2.0 |
| Input Resolution | 256x256 |
| GMACs | 2.5 |
| Dataset | ImageNet-1k |
What is eca_botnext26ts_256.c1_in1k?
ECA-BotNeXt26ts is an innovative image classification model that combines Efficient Channel Attention (ECA) with BotNet architecture, based on the ResNeXt backbone. Developed using timm's flexible BYOBNet framework, this model represents a careful balance between performance and computational efficiency, featuring 10.6M parameters and optimized for 256x256 image inputs.
Implementation Details
The model implements a hybrid architecture utilizing both convolutional neural networks and self-attention mechanisms. It's trained using SGD with Nesterov momentum and employs Adaptive Gradient Clipping (AGC). The training process follows the ResNet Strikes Back C recipes, incorporating a cosine learning rate schedule with warmup.
- Flexible block/stage layout with BYOBNet architecture
- Integrated stochastic depth and gradient checkpointing
- Layer-wise learning rate decay
- Per-stage feature extraction capabilities
Core Capabilities
- Image classification with state-of-the-art accuracy
- Feature map extraction at multiple scales
- Image embedding generation
- Support for both classification and backbone usage
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its efficient combination of ECA and BotNet architectures, offering a good balance between computational cost (2.5 GMACs) and model size (10.6M parameters). It's specifically tuned for reduced frequency of self-attention blocks while maintaining strong performance.
Q: What are the recommended use cases?
The model is particularly well-suited for image classification tasks, feature extraction, and as a backbone for downstream computer vision tasks. It's optimized for scenarios requiring 256x256 image inputs and where model efficiency is important.