| Property | Value |
|---|---|
| Parameter Count | 28.3M |
| Model Type | Image Classification |
| License | MIT |
| Paper | RepVGG: Making VGG-style ConvNets Great Again |
| Dataset | ImageNet-1k |
What is repvgg_a2.rvgg_in1k?
RepVGG A2 is a modern reimagining of the classic VGG architecture, designed to combine the simplicity of VGG-style networks with modern performance optimizations. This model implements 28.3M parameters and achieves efficient image classification through a carefully crafted architecture utilizing TIMM's flexible BYOBNet framework.
Implementation Details
The model is built using TIMM's BYOBNet (Bring-Your-Own-Blocks Network) implementation, allowing for highly configurable architecture components. It features 5.7 GMACs computational complexity and operates on 224x224 pixel images, producing 6.3M activations during processing.
- Flexible block and stage layout configuration
- Customizable stem layout and output stride
- Support for various activation and normalization layers
- Integrated channel and spatial/self-attention capabilities
Core Capabilities
- Image classification with ImageNet-1k dataset support
- Feature backbone extraction for downstream tasks
- Gradient checkpointing for memory efficiency
- Layer-wise learning rate decay
- Per-stage feature extraction capabilities
Frequently Asked Questions
Q: What makes this model unique?
RepVGG A2 stands out through its modern interpretation of the VGG architecture, offering simple inference-time architecture while maintaining competitive performance. It combines the benefits of skip connections during training while eliminating them during inference, resulting in a highly efficient deployment model.
Q: What are the recommended use cases?
This model is particularly well-suited for image classification tasks, feature extraction, and as a backbone for more complex computer vision applications. Its efficient architecture makes it especially suitable for production deployments where inference speed and resource efficiency are important.