| Property | Value |
|---|---|
| Parameter Count | 88.7M |
| Model Type | Image Classification / Feature Backbone |
| Input Resolution | 384 x 384 |
| Top-1 Accuracy | 87.646% |
| GMACs | 45.21 |
| Paper | ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders |
What is convnextv2_base.fcmae_ft_in22k_in1k_384?
This is a state-of-the-art convolutional neural network that represents the base variant of the ConvNeXt V2 architecture. It was pretrained using a fully convolutional masked autoencoder (FCMAE) framework and subsequently fine-tuned on ImageNet-22k and ImageNet-1k datasets. The model operates on 384x384 pixel images and achieves an impressive 87.646% top-1 accuracy.
Implementation Details
The model features a sophisticated architecture with 88.7M parameters and requires 45.2 GMACs (billion multiply-accumulate operations) for inference. It maintains 84.5M activations during processing and delivers efficient performance with 209.51 samples per second at a batch size of 256.
- Advanced FCMAE pretraining methodology
- Hierarchical feature extraction capabilities
- Optimized for 384x384 resolution inputs
- Dual-stage fine-tuning on ImageNet-22k and ImageNet-1k
Core Capabilities
- High-accuracy image classification
- Feature map extraction at multiple scales
- Image embedding generation
- Transfer learning applications
Frequently Asked Questions
Q: What makes this model unique?
This model combines the innovative ConvNeXt V2 architecture with FCMAE pretraining, offering an excellent balance between performance (87.646% top-1 accuracy) and efficiency (209.51 samples/sec). It's particularly notable for its ability to process high-resolution 384x384 images while maintaining strong performance.
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 high-resolution image processing and where both accuracy and efficiency are important considerations.