mean vs identity pooling?

[vr25]

Hi,

The paper describes four pooling functions: 1. Mean, 2. Identity, 3. Transformer, and 4. LSTM.

I am confused between mean and identity. I follow that mean means simply average all the [CLS] embeddings for all the chunks which would result in a final [768] -dimensional vector. In this way, how would identity function work? Does it mean concatenating all [CLS] vectors and if so, wouldn't it turn into a very long vector like: number of chunks x 768 ?

Any help in understanding this concept would be appreciated!

Thanks!

[AndriyMulyar]

Yes if you fix the number of chunks.

…

On Mon, Nov 2, 2020, 10:27 AM Vipula Rawte ***@***.***> wrote: Hi, The papers describe four pooling functions: 1. Mean, 2. Identity, 3. Transformer, and 4. LSTM. I am confused between mean and identity. I follow that mean means simply average all the [CLS] embeddings for all the chunks which would result in a final [768] -dimensional vector. In this way, how would identity function work? Does it concatenating all [CLS] vectors and if so, wouldn't it turn into a very long vector like: number of chunks x 768? Any help in understanding this concept would be appreciated! Thanks! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#15>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADJ4TBRJ6VIMEXNYMZYEOOTSN3FXJANCNFSM4THTLMKA> .

[vr25]

So in your case, it would be 20 x 768? (since max # chunks = 20)?

[n8henrie]

I'm also confused.

This paper references this repo as its source code and refers to the four models noted above: 1. Mean, 2. Identity, 3. Transformer, and 4. LSTM.

However, in the code, I see LSTM and Transformer, but instead of Mean and Identity, I see Linear and MaxPool: https://github.com/AndriyMulyar/bert_document_classification/blob/572883204cb1aca50d346979319905f698ad7049/bert_document_classification/document_bert_architectures.py

It looks like the Linear model matches the description of the Identity (reshaping output into a concatenation of CLS via bert_output.view(bert_output.shape[0], -1)).

But MaxPool looks to be doing something totally different than what is described: bert_output.max(dim=1)[0] vs A dimension-wise mean over all CLS embeddings.

@AndriyMulyar thank you very much for providing your code. I'm a novice here -- is there something I'm missing?

EDIT: I think it's also worth noting that the Mean model was the top performer in the paper, so it seems odd to have a different implementation here than what is described.

[AndriyMulyar]

Hi, the public codebase just hasn't been updated. You can change the pooling from max to mean in the implementation to replicate the stated results in the paper.

Cheers