EfficientNet B1 RA4
| Property | Value |
|---|---|
| Parameter Count | 7.86M |
| License | Apache 2.0 |
| Top-1 Accuracy | 81.44% |
| Image Size | 288x288 (test), 240x240 (train) |
| Paper | EfficientNet Paper |
What is efficientnet_b1.ra4_e3600_r240_in1k?
This is an optimized variant of EfficientNet B1 trained on ImageNet-1k, incorporating training improvements from timm. It represents a balanced trade-off between model size and accuracy, achieving 81.44% top-1 accuracy while maintaining a relatively small parameter count of 7.86M.
Implementation Details
The model utilizes the EfficientNet architecture with specific optimizations from the RA4 training regime, trained for 3600 epochs. It employs advanced training techniques inspired by MobileNet-V4 and "ResNet Strikes Back" methodologies.
- Optimized for both training (240x240) and inference (288x288) resolutions
- Implements efficient mobile-first architecture design
- Features balanced GMAC count of 0.7 and 10.9M activations
Core Capabilities
- Image classification with 1000 classes
- Feature extraction capabilities
- Efficient inference on mobile/edge devices
- Optimized for both accuracy and model size
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its excellent efficiency-to-accuracy ratio, achieving 81.44% top-1 accuracy with only 7.86M parameters, making it ideal for resource-constrained deployments while maintaining competitive performance.
Q: What are the recommended use cases?
The model is well-suited for mobile and edge device deployment, particularly for applications requiring high-quality image classification while maintaining efficient resource usage. It's especially effective for scenarios requiring a balance between model size and accuracy.
