| Property | Value |
|---|---|
| Parameter Count | 28.6M |
| GMACs | 4.5 |
| Training Image Size | 224 x 224 |
| Testing Image Size | 288 x 288 |
| Paper | A ConvNet for the 2020s |
What is convnext_tiny.in12k_ft_in1k?
This is a ConvNeXt tiny model that represents a modern approach to convolutional neural networks. Initially pretrained on ImageNet-12k (a subset of ImageNet-22k with 11,821 classes) and then fine-tuned on ImageNet-1k, this model achieves an impressive 84.186% top-1 accuracy at 224px resolution. The training was conducted on TPUs through the TRC program, with fine-tuning performed on 8x GPU Lambda Labs cloud instances.
Implementation Details
The model features a compact architecture with 28.6M parameters and requires 4.5 GMACs for inference. It processes images with a training resolution of 224x224 pixels and can handle 288x288 during testing. The architecture maintains efficient memory usage with 13.4M activations.
- Efficient parameter utilization with 28.6M parameters
- Optimized for both accuracy and speed with 4.5 GMACs
- Flexible resolution handling between training and inference
- Supports feature extraction and embedding generation
Core Capabilities
- Image classification with 1000 classes
- Feature map extraction at multiple scales
- Image embedding generation
- High throughput with 2433.7 samples/sec at batch size 256
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its efficient architecture that combines the benefits of modern ConvNet design with practical efficiency. Its pre-training on ImageNet-12k followed by fine-tuning on ImageNet-1k provides robust feature learning while maintaining a reasonable parameter count.
Q: What are the recommended use cases?
The model is well-suited for image classification tasks, feature extraction, and as a backbone for downstream computer vision tasks. It offers a good balance between accuracy and computational efficiency, making it suitable for both research and production environments.