Why generated bert embeddings are coming "nan", from the get_bert_embeddings function and the loss of model is also nan while training

[deepanshudashora]

I am training the mentioned https://github.com/jingxuanyang/Shopee-Product-Matching/blob/master/notebook-text/bert-base-multilingual-uncased-margin-0.8.ipynb method for training the model, but generated embeddings from bert while validation are nan and casing the training to stop

I am using a custom dataset of mine not the shopee dataset