This repo is a reimplementation of EfficientNet V2. The code base is heavily inspired by TensorFlow implementation and EfficientNet Keras
| Model | params | ImageNet Acc | CIFAR-10 |
|---|---|---|---|
| EfficientNetV2_Base | 7M | - | - |
| EfficientNetV2_S | 22M | 83.9 | 98.7 |
| EfficientNetV2_M | 54M | 85.1 | 99.0 |
| EfficientNetV2_L | 120M | 85.7 | 99.1 |
| EfficientNetV2_XL (21K) | 208M | 87.3 | - |
Ported from automl efficientnetv2 imagenet21k pretrained weights
| Arch | imagenet | imagenet21k | imagenet21k-ft1k |
|---|---|---|---|
| EfficientNetV2_S | h5 | h5 | h5 |
| EfficientNetV2_M | h5 | h5 | h5 |
| EfficientNetV2_L | h5 | h5 | h5 |
| EfficientNetV2_XL | - | h5 | h5 |
- Start the model and reload weights
from efficientnetv2 import EfficientNetV2_L
model = EfficientNetV2_L(input_shape=(512, 512, 3), weights=None)
model.load_weights('path_to_model.h5')- Reloading pretrained ImageNet21K weight
from efficientnetv2 import EfficientNetV2_S
model = EfficientNetV2_S(weights='imagenet21k-ft1k')- TensorFlow >= 2.4
pip install -U git+https://github.com/GdoongMathew/EfficientNetV2