I am using BERT (araBert to be more specific) for Arabic abstractive text summarization, but I don't want to train all the parameters from scratch. What I am looking for is a way to freeze the layers and then add some layers (LSTM or Transformer layers) to train the model on my data. How can I accomplish that? Because training the model from scratch takes a lot of resources and time. Thanks for your answers in advance

Tried the HuggingFace approach for fine-tuning summarization task, but it trains all the parameters.