Transformer - get contribution matrix

Hi everyone,

I am trying to experiment with Transformers. After training, I would like to visualize the contribution matrix from an input. If I understood correctly, the contribution matrix is a square one, indicating the contribution of each element with other elements from the sequence (diagonal values are high). This matrix is computed as in the linked picture, Q and K being respectively the query and key matrix, and d the number of features of each element.


Is there a way to easily get the result of this operation in order to visualize this contribution matrix?

I think that these contribution matrices can be obtained from the TransformerEncoderLayer. So, in the function at line 350 in this file: pytorch/ at master · pytorch/pytorch · GitHub, I call the function self.self_attn with parameter need_weights=True, average_attn_weights=False, and then save the output_weights as tensors. Is this the right way ?

Thanks for the help

Ok, thanks for the answer. So it seems it is the right way to do this.

Nevertheless, it seems like the obtained matrices don’t have diagonal values that are high, and there aren’t symmetric.

I think I miss something, is it possible to have more information on how to display those matrices?