About multi-head attention in attention is all you need, thanks. about transformer HOT 2 OPEN

As far as I understand, your doubt is that why Q, K, V is not going through n_head linear transformations to extract Q_i, Q_i and V_i corresponding to each head ?
=>
The answer to that is to avoid significant growth of computational cost and parametrization cost, we set d_q = d_k = d_v = d_model / n_head. [1]
This is what the split() function in MultiHeadAttention does, and then concat() function essentially is a weighted combination of these heads, just like in the paper.

You can see a similar implementation in PyTorch source code as well. [2]

Anyone else reading this, please correct me if I am wrong or if there are some others benefits/reasons of using this implementation.

EDIT:
It's clearly mentioned in the paper as well:

In this work we employ h = 8 parallel attention layers, or heads. For each of these we use d_k = d_v = d_model / h = 64. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.

For the most faithful implementations of research papers, you should also check out labml.ai annotated pytorch implementations repository. [3]

[1] https://d2l.ai/chapter_attention-mechanisms-and-transformers/multihead-attention.html
[2] https://pytorch.org/docs/stable/_modules/torch/nn/modules/activation.html#MultiheadAttention
[3] http://nlp.seas.harvard.edu/annotated-transformer/

from transformer.

As far as I understand, your doubt is that why Q, K, V is not going through n_head linear transformations to extract Q_i, Q_i and V_i corresponding to each head ? => The answer to that is to avoid significant growth of computational cost and parametrization cost, we set d_q = d_k = d_v = d_model / n_head. [1] This is what the split() function in MultiHeadAttention does, and then concat() function essentially is a weighted combination of these heads, just like in the paper.

You can see a similar implementation in PyTorch source code as well. [2]

Anyone else reading this, please correct me if I am wrong or if there are some others benefits/reasons of using this implementation.

EDIT: It's clearly mentioned in the paper as well:

In this work we employ h = 8 parallel attention layers, or heads. For each of these we use d_k = d_v = d_model / h = 64. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.

For the most faithful implementations of research papers, you should also check out labml.ai annotated pytorch implementations repository. [3]

[1] https://d2l.ai/chapter_attention-mechanisms-and-transformers/multihead-attention.html [2] https://pytorch.org/docs/stable/_modules/torch/nn/modules/activation.html#MultiheadAttention [3] http://nlp.seas.harvard.edu/annotated-transformer/

Thank you for your comment, but it doesn't address my question. For instance, consider a sequence, and we need to produce its embedding matrix, named X. Then, it is sent to every head and multiplied by W_q, W_k, and W_v, respectively. Now, each head generates its corresponding Q, K, and V.

However, before entering each linear layer in every head, the paper's multi-head attention illustration shows Q, K, and V instead of X, X, and X correspondingly.

from transformer.