The goal of this project is to make an alternative solution to the current keras version of the Bert encoder, publically available. In this version, we don't use hub layers. Instead we only use standard keras layers for easier model manipulation.

This implementation uses tf.keras.layers.EinsumDense which are part of release tensorflow 2.10+.

You can find different Bert model with tf.keras layers