SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Swin support for any input size (#15986)

Browse Source 
* padding done

* correctly return one attention per layer

* almost correct, attentions are not flatten one tuple per stage

* tests green

* doc

* conversations

* reshaping hidden_states

* view in the test

* reshape_hidden_states in Encoder and Model

* new outputs with reshaped_hidden_states

* conversations

* doc

* Update docs/source/model_doc/swin.mdx

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Apply suggestions from code review

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* conversations

* fix tests

* minor changes

* resolved conversations

* attentions one per stage

* typo

* typos

* typos

* function signature

* CI

* clean up tests

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
This commit is contained in:
Francesco Saverio Zuppichini
2022-03-16 18:38:25 +01:00
committed by GitHub
parent 204c54d411
commit 667b823b89
3 changed files with 317 additions and 118 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/model_doc/swin.mdx
View File

@@ -34,6 +34,8 @@ The hierarchical design and the shifted window approach also prove beneficial fo
Tips:
- One can use the [`AutoFeatureExtractor`] API to prepare images for the model.
- Swin pads the inputs supporting any input height and width (if divisible by `32`).
- Swin can be used as a *backbone*. When `output_hidden_states = True`, it will output both `hidden_states` and `reshaped_hidden_states`. The `reshaped_hidden_states` have a shape of `(batch, num_channels, height, width)` rather than `(batch_size, sequence_length, num_channels)`.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/swin_transformer_architecture.png"
alt="drawing" width="600"/>