Access the output of several layers of pretrained DistilBERT model

Question

I am trying to access the output embeddings from several different layers of the pretrained "DistilBERT" model. ("distilbert-base-uncased")

bert_output = model(input_ids, attention_mask=attention_mask)

The bert_output seems to return only the embedding values of the last layer for the input tokens.

Answer 1

If you want to get the output of all the hidden layers, you need to add the output_hidden_states=True kwarg to your config.

Your code will look something like

from transformers import DistilBertModel, DistilBertConfig

config = DistilBertConfig.from_pretrained('distilbert-base-cased', output_hidden_states=True)
model = DistilBertModel.from_pretrained('distilbert-base-cased', config=config)

The hidden layers will be made available as bert_output[2]