Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add support for CLS token pooling in text embedding #385

Merged
merged 6 commits into from
Aug 6, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion lib/bumblebee/text.ex
Original file line number Diff line number Diff line change
Expand Up @@ -313,7 +313,14 @@ defmodule Bumblebee.Text do
this option is ignored. Defaults to `:pooled_state`

* `:output_pool` - pooling to apply on top of the model output, in case
it is not already a pooled embedding. Supported values: `:mean_pooling`.
it is not already a pooled embedding. Supported values:

* `:mean_pooling` - performs a mean across all tokens

* `cls_token_pooling` - takes the embedding for the special CLS token.
Note that we currently assume that the CLS token is the first token
in the sequence

By default no pooling is applied

* `:embedding_processor` - a post-processing step to apply to the
Expand Down
21 changes: 10 additions & 11 deletions lib/bumblebee/text/text_embedding.ex
Original file line number Diff line number Diff line change
Expand Up @@ -56,22 +56,21 @@ defmodule Bumblebee.Text.TextEmbedding do
output
end

if output_pool != nil and Nx.rank(output) != 3 do
raise ArgumentError,
"expected the output tensor to have rank 3 to apply :output_pool, got: #{Nx.rank(output)}." <>
" You should either disable pooling or pick a different output using :output_attribute"
end

output =
case output_pool do
nil ->
output

:mean_pooling ->
case Nx.rank(output) do
3 ->
:ok

rank ->
raise ArgumentError,
"expected the output tensor to have rank 3 to apply :output_pool, got: #{rank}." <>
" You should either disable pooling or pick a different output using :output_attribute"
end
:cls_token_pooling ->
jonatanklosko marked this conversation as resolved.
Show resolved Hide resolved
Nx.take(output, 0, axis: 1)

:mean_pooling ->
input_mask_expanded = Nx.new_axis(inputs["attention_mask"], -1)

output
Expand All @@ -81,7 +80,7 @@ defmodule Bumblebee.Text.TextEmbedding do

other ->
raise ArgumentError,
"expected :output_pool to be one of nil or :mean_pooling, got: #{inspect(other)}"
"expected :output_pool to be one of :cls_token_pooling, :mean_pooling or nil, got: #{inspect(other)}"
end

output =
Expand Down
Loading