Question about the layer representation from the first token of the X #2
Description
Hello,
Thanks for the easy-read code. I have a question about the code, could you please help me?
In your paper, you said "we use the hidden state of the first token in text x" (between equations (6) and (7)). From my understanding, you are using the first token of the text x, which should be the [cls] token. However, in your code, representation, _ = pooling_layer(hidden_states) (line 730 in model.py). Looks like you are using the output from the layer attention network, which is in the size of [batch_size, seq_length, hidden_size], then put them into the pooling layer and get the layer representation in the size of [batch_size, hidden_size]. If so, could you please tell me why you choose this way, instead of the method in paper.
Please forgive me if I misunderstand this! And thanks for the great work!
Jiaxin