a$)(15$ points$)$ Take the pretrained BERT for BertForSequenceClassification model $.$ Extract logits for each token in the sentence using BERT for a sentence. So$,$ for example, if you have the sentence $$Hello$,$ how are you doing today?$,$ you will get $10$ vectors corresponding to each token in the sentence. To get the embeddings, you get the output of BERT $($output$)$ and do embeddings $=$ output.last$\_$hidden$\_$state.

b$)(15$ points$)$ Extract BERT embeddings for all positive and negative sentences in the train.tsv file we used in class for sentiment classification.

c$)(20$ points$)$ Train a single layer LSTM network with a hidden dimension of $128$ to do sentiment classification from BERT outputs. Use a learning rate of $0\mathrm{.}01,$ batch$-$size of $32.$

You don$$t have to tune the number of iterations. Just run a few iterations and report the accuracy.