| Property | Value |
|---|---|
| Parameter Count | 8.55M |
| Image Size | 240x240 |
| License | Apache 2.0 |
| Paper | CrossViT Paper |
| Dataset | ImageNet-1k |
What is crossvit_9_240.in1k?
CrossViT-9 is an innovative vision transformer model that implements a cross-attention multi-scale architecture for image classification. Developed by IBM researchers, this model represents a lightweight implementation with 8.55M parameters, specifically designed to process 240x240 pixel images while maintaining efficient computational requirements of just 1.8 GMACs.
Implementation Details
The model employs a unique dual-branch architecture that processes images at multiple scales simultaneously. It achieves this through cross-attention mechanisms that allow information exchange between different resolution pathways, resulting in more robust feature extraction.
- Multi-scale processing with cross-attention mechanism
- Efficient architecture with 8.6M parameters
- 9.5M activations for feature processing
- Optimized for 240x240 resolution inputs
Core Capabilities
- High-quality image classification on ImageNet-1k dataset
- Feature extraction for downstream tasks
- Efficient processing with reduced computational overhead
- Support for both classification and embedding generation
Frequently Asked Questions
Q: What makes this model unique?
CrossViT-9's distinctive feature is its cross-attention mechanism that enables effective multi-scale processing while maintaining a compact parameter count. This makes it particularly efficient for real-world applications where computational resources may be limited.
Q: What are the recommended use cases?
The model is well-suited for image classification tasks, particularly when working with fixed 240x240 resolution images. It can be used for both direct classification and as a feature extractor for transfer learning applications, with support for both full classification and embedding generation workflows.