Skip to content
GitLab
  1. 08 Jul, 2020 2 commits
  3. 07 Jul, 2020 3 commits
    • Yuxin Wu's avatar
      make ops in trainer more async · 2984c56c
      Yuxin Wu authored 
      Summary:
Both _write_metrics and _detect_anomaly require stream synchronization.
Therefore move them into a separate stream and after launching DDP allreduce.
Improves 1~2% throughput and the speed is similar to removing both.

Reviewed By: rbgirshick

Differential Revision: D22381820

fbshipit-source-id: 9628f90725f5daea49f04f7988fea954da236331
      2984c56c
    • Yuxin Wu's avatar
      use loss_weight=dict/float for box_head · 3a1d2fde
      Yuxin Wu authored 
      Summary: similar to RPN

Reviewed By: rbgirshick

Differential Revision: D22367132

fbshipit-source-id: 9d3dfced6943d291ca18c3d46a126d93fc06b796
      3a1d2fde
    • Davi Nakajima An's avatar
      Added Photometric Transform class to Detectron2 · 84b183bc
      Davi Nakajima An authored 
      Summary: Created a generic PhotometricTransform for transformations that only affect the color aspect of the image and not the coordinates

Reviewed By: ppwwyyxx

Differential Revision: D22301246

fbshipit-source-id: 06e713333ac9ff44ac0f0c564ada33c1f9844815
      84b183bc
  5. 04 Jul, 2020 1 commit
  7. 03 Jul, 2020 3 commits
  9. 30 Jun, 2020 2 commits
  11. 29 Jun, 2020 2 commits
  13. 28 Jun, 2020 1 commit
    • Jon Morton's avatar
      Configurable loss for rpn box regression and giou support · af0462a8
      Jon Morton authored 
      Summary:
Add support for using giou loss for rpn proposal regression.
Add options for configuring rpn bbox reg loss type and loss weight.

Reviewed By: ppwwyyxx

Differential Revision: D16974627

fbshipit-source-id: 43e0cba64a1ac403ada69955755de66285c9631c
      af0462a8
  15. 27 Jun, 2020 2 commits
    • Yuxin Wu's avatar
      add __all__ in detection_utils · e45cd680
      Yuxin Wu authored 
      Summary: don't export everything

Reviewed By: rbgirshick

Differential Revision: D22262954

fbshipit-source-id: dbf82221fbc320525645add4febe44d1cd78a74b
      e45cd680
    • Zach Carmichael's avatar
      Correctly apply EXIF orientation to images · 75ebc534
      Zach Carmichael authored 
      Summary:
Workaround for the PIL bug https://github.com/python-pillow/Pillow/issues/3973 - the exif orientation metadata is used to correctly transpose a PIL image when reading an image. The current approach silences any exceptions here, leading to strange behavior, such as issue https://github.com/facebookresearch/detectron2/issues/933. This is indeed a PIL issue, but the fix here performs the same function that is necessary to correctly orient images. The only differences between the committed function `apply_exif_orientation`  and the PIL `exif_transpose` is that a copy of the image is not generated if no transpose is needed, and that the exif data is not updated in the transposed image. As `read_image` returns a NumPy ndarray, this information is discarded, anyway.
Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1666

Differential Revision: D22238978

Pulled By: ppwwyyxx

fbshipit-source-id: 497c97b99fbfcf9857cbdf7516df10da4b42278d
      75ebc534
  17. 26 Jun, 2020 1 commit
    • Vasil Khalidov's avatar
      correct transform data loading · aaae87c9
      Vasil Khalidov authored 
      Summary:
Correct transform data loading to satisfy the following requirements:
1. Train-time data loaders that use the transforms and are executed in separate processes should operate on CPU only;
2. Test-time augmentation should be able to convert the transform tensors to a particular device

Reviewed By: patricklabatut

Differential Revision: D22066573

fbshipit-source-id: e2b97c362e05d07f734987ba768db4b521f3c897
      aaae87c9
  19. 25 Jun, 2020 5 commits
    • Yuxin Wu's avatar
      TransformGen->Augmentation, part 2 · c0f0e1d0
      Yuxin Wu authored 
      Summary: Rename the file and update import paths.

Reviewed By: rbgirshick

Differential Revision: D22201091

fbshipit-source-id: 10ca2442732b434f6f2eebfc81631df69fc3f08c
      c0f0e1d0
    • Vasil Khalidov's avatar
      add inference-based loader · a326cb8c
      Vasil Khalidov authored 
      Summary:
Adds an inference-based loader and associated tests.
Inference-based loader loader consists of:
 - an input data loader
 - a model
 - a data sampler
Inference-based loader takes input data batches from the input data loader, applies the model to each batch and outputs data produced by the data sampler when applied to model outputs.

Reviewed By: patricklabatut

Differential Revision: D21881798

fbshipit-source-id: 7614ab316cc7acff6141dde65a2c64495ca72da2
      a326cb8c
    • Vasil Khalidov's avatar
      adjust video keyframe dataset · e389a5c9
      Vasil Khalidov authored 
      Summary:
Adjust the interface of `VideoKeyframeDataset`:
1. Can be imported from `densepose.data.video`, not from `densepose.data`;
2. `list_keyframes` and `read_keyframes` now both accept path-like object + stream index;
3. Fix a bug with arguments order in `video_list_from_file`

Reviewed By: patricklabatut

Differential Revision: D22153239

fbshipit-source-id: 37a358dfbe8270368fb16fae79a56f17c8a5d573
      e389a5c9
    • chenbohua3's avatar
      Make anchor_generator scriptable · d8290953
      chenbohua3 authored 
      Summary:
This is a refined version of https://github.com/facebookresearch/detectron2/issues/1483  to make anchor_generator scriptable. Thanks for your work ppwwyyxx :)
Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1518

Reviewed By: rbgirshick

Differential Revision: D22213759

Pulled By: ppwwyyxx

fbshipit-source-id: 5796548e4cd33dedd3d09de0f8ab03ea032bb7c9
      d8290953
    • Yuxin Wu's avatar
      TransformGen -> Augmentation · c71fbd8c
      Yuxin Wu authored 
      Reviewed By: rbgirshick

Differential Revision: D22177876

fbshipit-source-id: 6e9a04dc1738c6056a95a50534093f1e8481f4f9
      c71fbd8c
  21. 24 Jun, 2020 5 commits
  23. 23 Jun, 2020 3 commits
  25. 21 Jun, 2020 1 commit
    • RUPESH's avatar
      Added an import line · 0de7e8c6
      RUPESH authored 
      Summary:
The current method had an import issue for DensePoseResult hence added a line to resolve.

This PR is just an addition to the documentation and hence won't cause any conflicts.

Thanks for your contribution!

If you're sending a large PR (e.g., >50 lines),
please open an issue first about the feature / bug, and indicate how you want to contribute.

Before submitting a PR, please run `dev/linter.sh` to lint the code.

See https://detectron2.readthedocs.io/notes/contributing.html#pull-requests
about how we handle PRs.
Pull Request resolved: https://github.com/facebookresearch/detectron2/pull/1644

Differential Revision: D22152589

Pulled By: ppwwyyxx

fbshipit-source-id: 8aad10aec2083733ed2e90815db75c5ecf95890e
      0de7e8c6
  27. 20 Jun, 2020 2 commits
  29. 19 Jun, 2020 3 commits
  31. 18 Jun, 2020 2 commits
  33. 17 Jun, 2020 1 commit
    • Yuxin Wu's avatar
      use pycocotools from pypi · ae931b47
      Yuxin Wu authored 
      Summary: it's updated!

Reviewed By: rbgirshick

Differential Revision: D22076151

fbshipit-source-id: a2923e3bd247d7208941c8ec49178ab42db5af32
      ae931b47
  35. 16 Jun, 2020 1 commit