Why do you use "self.seq" (which means two Bilstm) before transeformer in model/__init.py ?

[yoyoshuang]

Thanks for you great job.

I am not very understand this part "Why do you use "self.seq" (which means two Bilstm) before transeformer in model/__init.py ?"

What is the net learning in this part?

Would you please help me to understand this part?

And in my work , I have try the same way to combine the CNN's feature map to the transeformer part , but the results looks like accoured something wrong, after adding this "self.seq" part, the work seems right , I can't explain it .

Thank you for you helping!!!

[opconty]

Hi, Thanks for your attention, according to the paper https://arxiv.org/abs/1904.01906,

 self.seq = nn.Sequential(BiLSTM(output_channel, hidden_size, hidden_size),
                                 BiLSTM(hidden_size, hidden_size, hidden_size))

is the sequence module, bilstm is not necessary and you can remove it,