I created a repository with EfficientNet-lite model variants adapted to Keras (functional API).
The main goal was to mimic
tf.keras.applications usability as much as possible. The lite model variants were only available in Tensorflow 1.x - adopting them to Keras allowed for more flexibility, and made them more consistent with existing API and documentation.
According to original repository, the lite variants:
- Use ReLU6 instead of Swish.
- Do not use SE blocks.
- Have fixed Stem and Head, while scaling up the models.
Hope that it helps somebody!