tf_efficientnetv2_m.in21k
| Property | Value |
|---|---|
| Parameters | 80.8M |
| GMACs | 15.9 |
| Training Image Size | 384x384 |
| Test Image Size | 480x480 |
| Dataset | ImageNet-21k |
| Paper | EfficientNetV2: Smaller Models and Faster Training |
What is tf_efficientnetv2_m.in21k?
This is a PyTorch port of the TensorFlow EfficientNetV2-M model, originally trained on ImageNet-21k by the paper authors and converted by Ross Wightman. It represents a significant advancement in efficient image classification, balancing model size and computational requirements while maintaining high accuracy.
Implementation Details
The model features 80.8M parameters and requires 15.9 GMACs for inference, with 57.5M activations. It's designed to process images at 384x384 resolution during training and 480x480 during testing, showcasing its ability to handle different input scales effectively.
- Optimized architecture for improved training efficiency
- Supports feature extraction with multiple output scales
- Capable of generating image embeddings
- Includes model-specific preprocessing transforms
Core Capabilities
- Image classification with ImageNet-21k classes
- Feature map extraction at multiple scales
- Dense embedding generation
- Flexible input resolution handling
Frequently Asked Questions
Q: What makes this model unique?
This model represents a balanced approach to efficient deep learning, offering state-of-the-art performance while maintaining reasonable computational requirements. Its training on ImageNet-21k provides rich feature representations for transfer learning tasks.
Q: What are the recommended use cases?
The model is well-suited for image classification tasks, feature extraction for downstream tasks, and generating image embeddings for similarity-based applications. It's particularly valuable when working with large-scale image classification problems or as a backbone for transfer learning.
