How to concatenate additional features to the last layer of Bert

🤗Transformers
1

Hi
I have sentences with additional features for each (a 5 dims vector of floats) and a label for each (True or False).
So an example of sample:
“hello have a nice day”, [0.2, 0.1, 0.6, 0.7, 0.2], True
I have a classification task on my data.
I want to finetune Bert to take input_sentence, concat the vec to the last hidden layer, and predict label_sentence.
How can it be done? I dont find any code sample of concat additional data to the last layer before classification

Thanks!

2

One way you could do it is by precomputing the last hidden states’ CLS token embedding for each of the text in your dataset and storing it in a numpy array. Then you could concatenate this array with your desired additional features to accomplish classification task.

On a side note, you may want to rescale your additional features to the scale of bert embeddings.

3

Hi,
I want it to be as a wrapping class to BERT so I will be able to use it in trainer, etc

4

Hi,

This question has been answered here: Combine BertForSequenceClassificaion with Additional Features - #3 by nielsr

5

@nielsr Thanks, it is working great, just a few questions about ur notebook:

  1. In CustomSequenceClassification, why do you need to call post_init()?
  2. Lets say that in the input, instead of a single sentence I want to insert two sentences (“This is a sentence 1”, “This is a sentence 2”) , how would you do it?

Also, if you could please help with an additional question I just posted it would be highly appreciated…

Thanks

6

Hi,

  1. In CustomSequenceClassification, why do you need to call post_init()?

The post_init method takes care of initializing all the weights as defined in the _init_weights method of the xxxPreTrainedModel. See e.g. the _init_weights method of BERT here.

  1. Lets say that in the input, instead of a single sentence I want to insert two sentences (“This is a sentence 1”, “This is a sentence 2”) , how would you do it?

You can leverage the tokenizer for that, as it supports pair of sentences besides single sentences. Just like this:

inputs = tokenizer(sentence_1, sentence_2, padding=True, return_tensors="pt")

This will then look like [CLS] sentence_1 [SEP] sentence_2 [SEP] in terms of tokens.

7
  1. If the line of

init_weights()

would been remove, what will be the effect? This init the weight of the classification head? are they not automatically initilized?

  1. Do should I change to:

ds = Dataset.from_dict({“text”: [[“This is a sentence1”, “This is a sentence2” ]]*100, “extra_data”: np.random.randint(0, 10, size=(100, 5)), “labels”: np.random.randint(0, 2, size=(100,))})
tokenized_ds = ds.map(lambda x: tokenizer(x[“text”][0], x[“text”][1]))
?