How to extract attention weights about asformer HOT 4 OPEN

Sorry for my late response. You can save the attention weights during testing

from asformer.

ok, thanks for the reply.
Just to be sure the attention weight are the 'attention' matrix of the following line of code, right?
output, attention = self.att_helper.scalar_dot_att(q, k, v, final_mask)

In my case I extracted them from the '_sliding_window_self_att(self, q, k, v, mask)' function.
Thanks

from asformer.

I am asking because I find that most of the time if an encoder find a strong attention to some frames inside his windows, the enoder of the next block will give 0 (red cross) attention for the same frames. As you can see from the image below. So maybe I am extract the attention weight in the wrong way. This is the code I use to extract them:
att_weight = attention.reshape([attention.shape[0] * attention.shape[2], attention.shape[1]])
save_dir_att = 'attention_weight'
torch.save(att_weight, save_dir_att + '/att_weigth_window_size_' + str("%03d" % (b.shape[1],)) + '.t')
the code is inside the '_sliding_window_self_att(self, q, k, v, mask)' function

from asformer.

Sorry to bother you again, but there is a reason behind it or is it just that I am extracting the attention weights in the wrong way? Thanks for your time

from asformer.