RegNetY-120 SW ImageNet-12k Fine-tuned Model
| Property | Value |
|---|---|
| Parameters | 51.8M |
| GMACs | 12.1 |
| Training Image Size | 224x224 |
| Testing Image Size | 288x288 |
| Paper | Designing Network Design Spaces |
What is regnety_120.sw_in12k_ft_in1k?
This is a RegNetY architecture model with 51.8M parameters, initially pretrained on ImageNet-12k and then fine-tuned on ImageNet-1k. It represents a significant advancement in efficient neural network design, achieving 85.4% top-1 accuracy at 288px resolution. The model implements the RegNet architecture principles with additional enhancements from the timm library.
Implementation Details
The model incorporates several advanced features implemented in the timm framework, making it more powerful than standard RegNet implementations. It utilizes a sophisticated architecture that balances computational efficiency with model performance.
- Stochastic depth for improved regularization
- Gradient checkpointing for memory efficiency
- Layer-wise learning rate decay
- Configurable output stride with dilation support
- Flexible activation and normalization layers
Core Capabilities
- Image classification with 1000 classes
- Feature extraction with multiple resolution outputs
- Embedding generation for downstream tasks
- Efficient inference with various input resolutions
Frequently Asked Questions
Q: What makes this model unique?
This model stands out due to its unique combination of ImageNet-12k pretraining and ImageNet-1k fine-tuning, along with timm-specific enhancements that improve both training and inference performance. It offers an excellent balance between accuracy (85.4% top-1) and computational efficiency (12.1 GMACs).
Q: What are the recommended use cases?
The model is particularly well-suited for production image classification tasks, feature extraction for downstream computer vision applications, and as a backbone for transfer learning. It performs optimally with 288x288 pixel images during inference, though it can handle 224x224 with slightly reduced accuracy.
