EfficientNet B3 RA2
| Property | Value |
|---|---|
| Parameter Count | 12.3M |
| License | Apache-2.0 |
| Architecture | EfficientNet B3 with RandAugment |
| Training Dataset | ImageNet-1k |
| Image Size | Train: 288x288, Test: 320x320 |
What is efficientnet_b3.ra2_in1k?
The efficientnet_b3.ra2_in1k is an advanced implementation of the EfficientNet architecture, specifically the B3 variant, trained using the RandAugment (RA2) recipe. This model represents a careful balance between computational efficiency and accuracy, featuring 12.3M parameters and 1.6 GMACs.
Implementation Details
This model utilizes the RandAugment 'RA2' recipe, which evolved from the original EfficientNet augmentation strategies and was published in the "ResNet Strikes Back" paper. It employs RMSProp optimization with TF 1.0 behavior and implements EMA weight averaging for improved stability.
- Advanced augmentation strategy using RandAugment (RA2)
- Step-based learning rate schedule with warmup
- Optimized for both training (288x288) and inference (320x320)
- 21.5M activations with 1.6 GMACs computational requirement
Core Capabilities
- Image Classification with state-of-the-art performance
- Feature extraction for downstream tasks
- Embedding generation for similarity tasks
- Multi-scale feature map extraction
Frequently Asked Questions
Q: What makes this model unique?
This model stands out due to its implementation of the RA2 recipe, which provides enhanced training efficiency and performance compared to standard EfficientNet models. The careful balance of model size and computational requirements makes it particularly suitable for production deployments.
Q: What are the recommended use cases?
The model excels in image classification tasks, feature extraction, and as a backbone for transfer learning. It's particularly well-suited for applications requiring a good balance between accuracy and computational efficiency.
