Skip to content

D
detectron
Switch branch/tag
  1. 19 Sep, 2018 1 commit
  3. 15 Sep, 2018 1 commit
  5. 06 Sep, 2018 1 commit
    • Ashwin Bharambe's avatar
      Fix a bug in RPN proposal label generation · 3a38b7bc
      Ashwin Bharambe authored 
      Summary:
Negative labels were never set if `len(bg_inds) <= num_bg`. This
condition happens when there aren't as many negative anchor samples compared to the
typical batch size (256). Since detection scales are typically ~800, there are
plenty of negative samples and this bug is likely never hit. In any case, it
is certainly a bug which should be fixed!

Reviewed By: rbgirshick

Differential Revision: D9647703

fbshipit-source-id: fada9f4f2e0a0b5c1619940328c505e73f9a1c7e
      3a38b7bc
  7. 03 Sep, 2018 1 commit
  9. 31 Aug, 2018 1 commit
  11. 29 Aug, 2018 1 commit
  13. 24 Aug, 2018 1 commit
    • Ross Girshick's avatar
      Make data loader threads daemonic · c9ed587c
      Ross Girshick authored 
      Summary: Make data loader threads daemonic so the main process can exit when an exception is raised.

Reviewed By: ashwinb, newstzpz

Differential Revision: D9478059

fbshipit-source-id: 00bfaab51295b63c2f95ac1be99ae56a0ccf9d94
      c9ed587c
  15. 21 Aug, 2018 1 commit
    • Ross Girshick's avatar
      Avoid negatives areas found exception · 47e457a5
      Ross Girshick authored 
      Summary:
In some cases, eg training from scratch with a high LR or w/o activation normalization, we saw a "Negative Areas Found" exception. The RPN proposal generation code filters proposals that have a height or width that are too small. By default we use a `min_size` of 0. In a distant other part of the code, it computes the area of the proposals in order to determine FPN scale assignment. An issue arises because the `+ 1` in the `x2 - x1 + 1` width (and height) computation is not scale invariant. It turns out that the filter based on `min_size` was performed in the scaled input coordinates while the FPN scale assignment areas are computed using the unscaled, original image size. Thus a `min_size` of 0 can result in proposals with a negative area as measured wrt the original image. Uuuugh :P.

This diff addresses the issue by filtering based on the area in the unscaled input image and also using a generous (though still small) margin of 1 pixel for `min_size`.

Reviewed By: KaimingHe

Differential Revision: D8773216

fbshipit-source-id: b4ffbc5b6a831176b2656810edea3d3d4e52d687
      47e457a5
  17. 17 Aug, 2018 1 commit
  19. 08 Aug, 2018 1 commit
  21. 03 Aug, 2018 1 commit
  23. 02 Aug, 2018 1 commit
    • Ashwin Bharambe's avatar
      Set logging level correctly in utils/net.py · 40c39925
      Ashwin Bharambe authored 
      Summary:
AFAIK, this is what the rest of the codebase does. I don't want to
see logger.debug() messages show up during normal runs.

Differential Revision: D9071531

fbshipit-source-id: 4ac5ef57ee47bf958e35d723c379f695554f0ef0
      40c39925
  25. 31 Jul, 2018 1 commit
  27. 27 Jul, 2018 1 commit
  29. 14 Jul, 2018 1 commit
  31. 04 Jul, 2018 1 commit
    • Kaiming He's avatar
      Allow copy TRAIN.WEIGHTS to checkpoint folder · b2d24147
      Kaiming He authored 
      Summary:
Add TRAIN.COPY_WEIGHTS (default False):

When setting True, training will copy TRAIN.WEIGHTS to checkpoint folder and treat it as a candidate checkpoint. This is useful if we want to resume a checkpoint from another job (instead of using ImageNet pre-training).

Reviewed By: rbgirshick

Differential Revision: D8710711

fbshipit-source-id: 9c011227867c52eaaeb40e3c2679cff1cfccdc66
      b2d24147
  33. 03 Jul, 2018 3 commits
  35. 28 Jun, 2018 1 commit
  37. 26 Jun, 2018 1 commit
  39. 22 Jun, 2018 1 commit
  41. 21 Jun, 2018 1 commit
    • Norman Mu's avatar
      Fix typo in detectron config · ea9c76d9
      Norman Mu authored 
      Summary: change _DEPCRECATED_KEYS -> _DEPRECATED_KEYS

Differential Revision: D8553498

fbshipit-source-id: 1431a9552efce47be3aa40718c954521481950de
      ea9c76d9
  43. 12 Jun, 2018 2 commits
    • Kaiming He's avatar
      Update from-scratch results in GN/README.md · b5dada9b
      Kaiming He authored 
      Summary: Update from-scratch results in GN/README.md, given the freeze_at=0 fix.

Reviewed By: ir413

Differential Revision: D8369007

fbshipit-source-id: a7df13364125dac09379c5aed647f4f9a8bbf23b
      b5dada9b
    • Kaiming He's avatar
      Move freeze_at to config; fix an issue of GN from-scratch · f8ffc87c
      Kaiming He authored 
      Summary:
Move freeze_at to config.

Fix GN from-scratch scripts with freeze_at=0.

Reviewed By: rbgirshick

Differential Revision: D8325882

fbshipit-source-id: 9f362cb269ffd962056dfe8adc888083cff3d9b2
      f8ffc87c
  45. 11 Jun, 2018 1 commit
  47. 07 Jun, 2018 1 commit
    • Xinlei Chen's avatar
      Remove gt_boxes, seems no longer used? · a04ed18a
      Xinlei Chen authored 
      Summary: Remove code that seems redundant.

Reviewed By: rbgirshick

Differential Revision: D8200163

fbshipit-source-id: c3096c77a1c6b33eab5dc0ceb9414f3bdbbf71e9
      a04ed18a
  49. 06 Jun, 2018 1 commit
  51. 01 Jun, 2018 1 commit
  53. 26 May, 2018 1 commit
    • Kaiming He's avatar
      fix a bug of aspect ratio grouping · 0375b05e
      Kaiming He authored 
      Summary: Fix a bug of aspect ratio grouping to support training batch size > 2.

Reviewed By: rbgirshick

Differential Revision: D8157964

fbshipit-source-id: 28eec6246cd6daf83aef30da245fe7bf340fbbc9
      0375b05e
  55. 25 May, 2018 1 commit
  57. 24 May, 2018 5 commits
  59. 14 May, 2018 1 commit
  61. 09 May, 2018 1 commit
  63. 08 May, 2018 2 commits