Skip to content

GitLab

Switch branch/tag
  1. 21 Nov, 2022 5 commits
  3. 18 Nov, 2022 4 commits
    • Ali Hassani's avatar
      Add Neighborhood Attention Transformer (NAT) and Dilated NAT (DiNAT) models (#20219) · fc4a993e
      Ali Hassani authored 
      * Add DiNAT

* Adds DiNAT + tests

* Minor fixes

* Added HF model

* Add natten to dependencies.

* Cleanup

* Minor fixup

* Reformat

* Optional NATTEN import.

* Reformat & add doc to _toctree

* Reformat (finally)

* Dummy objects for DiNAT

* Add NAT + minor changes

Adds NAT as its own independent model + docs, tests
Adds NATTEN to ext deps to ensure ci picks it up.

* Remove natten from `all` and `dev-torch` deps, add manual pip install to ci tests

* Minor fixes.

* Fix READMEs.

* Requested changes to docs + minor fixes.

* Requested changes.

* Add NAT/DiNAT tests to layoutlm_job

* Correction to Dinat doc.

* Requested changes.
      fc4a993e
    • Nicolas Patry's avatar
      [Proposal] Breaking change `zero-shot-object-detection` for improved consistency. (#20280) · 8e777b3b
      Nicolas Patry authored 
      * [Proposal] Breaking change `zero-shot-object-detection` for improved
consistency.

This is a proposal to modify the output of `zero-shot-object-detection`
to provide better alignment with other pipelines.

The output is now strictly the same as `object-detection` whereas before
it would output lists of lists.

The name `candidate_labels` is used throughout for consistency with
other `zero-shot` pipelines.

The pipeline is changed to `ChunkPipeline` to support batching cleanly.

This removes all the lists and list of lists shenanigans, it's now a
matter of the base pipeline handling all this not this specific one.

**Breaking change**: It did remove complex calls potentials `pipe(images = [image1, image2],
text_queries=[candidates1, candidates2])` to support only
`pipe([{"image": image1, "candidate_labels": candidates1}, {"image": image2, "candidate_labels": candidates2}])`
when dealing with lists and/or datasets.
We could keep them, but it will add a lot of complexity to the code
base, since the pipeline is rather young, I'd rather break to keep the
code simpler, but we can revert this.

**Breaking change**: The name of the argument is now `image` instead of
`images` since it expects by default only 1 image. This is revertable
like the previous one.

**Breaking change**: The types is now simplified and flattened:

`pipe(inputs) == [{**object1}, {**object2}]`
instead of the previous
`pipe(inputs) == [[{**object1}, {**object1}], [{**object2}]]`
Where the different instances would be grouped by candidate labels
within lists.
IMHO this is not really desirable, since it would output empty lists and
is only adding superflous indirection compared to
`zero-shot-object-detection`.

It is relatively change free in terms of how the results, it does change
computation however since now the batching is handled by the pipeline
itself. It **did** change the results for the small models so there
seems to be a real difference in how the models handle this.

* Fixing the doctests.

* Behind is_torch_available.
      8e777b3b
    • atturaioe's avatar
      Add AnyPrecisionAdamW optimizer (#18961) · 84c9cc6d
      atturaioe authored 
      * Add AnyPrecisionAdamW optimizer

* Add optim_args argument to TrainingArgs

* Add tests for AnyPrecisionOptimizer

* Change AnyPrecisionAdam default params to float32

* Move default_anyprecision_kwargs in trainer test

* Rename AnyPrecisionAdamW
      84c9cc6d
    • amyeroberts's avatar
      Add padding image transformation (#19838) · b9826942
      amyeroberts authored 
      * Add padding transformation

* Add in upstream changes

* Update tests & docs

* Code formatting tuples in docstring
      b9826942
  5. 17 Nov, 2022 5 commits
  7. 16 Nov, 2022 2 commits
  9. 15 Nov, 2022 10 commits
  11. 14 Nov, 2022 6 commits
  13. 11 Nov, 2022 2 commits
  15. 10 Nov, 2022 4 commits
  17. 09 Nov, 2022 2 commits