Commits

Breadcrumbs

History for

ononnx

User selector

All users

Datepicker

All time

Commit History

Commits on Apr 6, 2019

• Added a keep_latest_interval setting to remove the need for a hard-coded 100k.

  dbolya

  committed
  f80441a

  Copy full SHA for f80441a

Commits on Mar 14, 2019

• Added the final base config and made train.py using keep_latest also keep every 100k th checkpoint for peace of mind.

  dbolya

  committed
  05c2fb8

  Copy full SHA for 05c2fb8

Commits on Mar 5, 2019

• Added a keep_latest argument to train.py that deletes the previous latest save after saving.

  dbolya

  committed
  eb9d528

  Copy full SHA for eb9d528

Commits on Mar 2, 2019

• Added *all* the loss functions.

  dbolya

  committed
  364d041

  Copy full SHA for 364d041

Commits on Feb 26, 2019

• Rewrote coefficient diversity loss to be much better and changed the way multibox_loss handles each induvial loss and how train prints them out. Now instead of returning 3 losses and adding togethe…

  Show description for a4052b3
  dbolya

  committed
  a4052b3

  Copy full SHA for a4052b3

Commits on Feb 22, 2019

• Implemented lr warmup.

  dbolya

  committed
  bf54385

  Copy full SHA for bf54385

• Refactored the way losses are printed during training so we can add an arbitrary number of additional losses with minimal extra effort.

  dbolya

  committed
  e61fa9d

  Copy full SHA for e61fa9d

• Removed the vizdom stuff to clean up train.py

  dbolya

  committed
  df6aef5

  Copy full SHA for df6aef5

Commits on Feb 15, 2019

• Added reset functionality to MovingAverage and reset them when the delayed config triggers.

  dbolya

  committed
  5824a85

  Copy full SHA for 5824a85

Commits on Feb 12, 2019

• Changed the default batch size to be 8 because why did we keep it as 32 for so long?

  dbolya

  committed
  72a4046

  Copy full SHA for 72a4046

Commits on Feb 10, 2019

• Did a fair bit of refactoring and changed the way we load datasets. I've included and this enables use to evaluate on the test set.

  dbolya

  committed
  d38bad1

  Copy full SHA for d38bad1

• Added the ability to specify the desired dataset in commandline.

  dbolya

  committed
  3340868

  Copy full SHA for 3340868

Commits on Feb 8, 2019

• Oh so that was why.

  dbolya

  committed
  f2f9eb1

  Copy full SHA for f2f9eb1

• Added a commandline argument to disable while I figure out how to get it to work with DataParallel.

  dbolya

  committed
  99cb6b2

  Copy full SHA for 99cb6b2

Commits on Feb 1, 2019

• Added a config setting that lets you specify certain iterations after which to dynamically change the config (i.e. delay turning on config settings). Yay Python!

  dbolya

  committed
  37068db

  Copy full SHA for 37068db

Commits on Oct 27, 2018

• Changed the combination of losses from individual gpus to be mean instead of sum since each gpu gets a smaller number of positives.

  dbolya

  committed
  3d30450

  Copy full SHA for 3d30450

Commits on Oct 25, 2018

• Forgot to cast tensors to the proper device.

  dbolya

  committed
  7e76e61

  Copy full SHA for 7e76e61

• Hopefully fixed the dataparallel scattering issue, hopefully, by adding a scatter wrapper class.

  dbolya

  committed
  8ab3f78

  Copy full SHA for 8ab3f78

Commits on Oct 11, 2018

• Added the ability to specify lr, momentum, weight decay, and gamma in a config.

  dbolya

  committed
  80478f4

  Copy full SHA for 80478f4

Commits on Oct 6, 2018

• Extended config 25_b to train for 800k iterations to see what would happen.

  dbolya

  committed
  3c2afad

  Copy full SHA for 3c2afad

Commits on Oct 2, 2018

• Finally added the ability to set a crowd iou threshold to count negatives that have a higher crowd iou than that to be considered neutral. I added 2 configs to test it and whether or not I borked a…

  Show description for 367c3e2
  dbolya

  committed
  367c3e2

  Copy full SHA for 367c3e2

Commits on Sep 29, 2018

• Added a line of text when calculating validation mAP to make it possible to differentiate freeqing from validation.

  dbolya

  committed
  451a764

  Copy full SHA for 451a764

Commits on Sep 28, 2018

• Oh... I figured out what the problem was. If you use --resume=latest and no latest file is found, the network starts from iteration 0, as intended. However, the way I implemented it, that branch ac…

  Show description for 898babd
  dbolya

  committed
  898babd

  Copy full SHA for 898babd

Commits on Sep 25, 2018

• Added the ability to specify "latest" for the train.py --resume option that will find the latest model for the given config.

  dbolya

  committed
  c3f7739

  Copy full SHA for c3f7739

Commits on Sep 14, 2018

• Fixed a crash when training past the last entry in step_iter.

  dbolya

  committed
  36a788e

  Copy full SHA for 36a788e

• Some refactoring to the make the code nicer and to remove some useless stuff. Also wrapped the loss function in a DataParallel, which should make training on multiple GPUs actually feasible.

  dbolya

  committed
  f89f901

  Copy full SHA for f89f901

Commits on Sep 8, 2018

• Fixed learning rate decay on resume such that lr decay is adjusted for any iterations after the lr_step, not when your iteration exactly equals the lr_step.

  dbolya

  committed
  435b3a6

  Copy full SHA for 435b3a6

Commits on Sep 5, 2018

• Change the way matching works so now you can specify a positive IoU minimum and a negative IoU maximum.

  dbolya

  committed
  e449871

  Copy full SHA for e449871

Commits on Aug 30, 2018

• Fixed a bug in validation because I wasn't giving a yolact object.

  dbolya

  committed
  d15b298

  Copy full SHA for d15b298

Commits on Aug 25, 2018

• Added the ability to specify a delay for starting prediction matching.

  dbolya

  committed
  b25479a

  Copy full SHA for b25479a

Commits on Aug 22, 2018

• Added no_bar.

  dbolya

  committed
  36eba64

  Copy full SHA for 36eba64

Commits on Aug 18, 2018

• Finally added validation mAP calculation to the training script. This will fire an evaluate call after every second epoch. Also made the mAP printing prettier.

  dbolya

  committed
  3f2ff36

  Copy full SHA for 3f2ff36

Commits on Aug 10, 2018

• Figured out why the ETA's weren't accurate: they didn't include data loading! Changed the timing system in train.py to be more accurate and not weird.

  dbolya

  committed
  d5a84c8

  Copy full SHA for d5a84c8

• Hopefully this fixes the vram issue.

  dbolya

  committed
  8750e23

  Copy full SHA for 8750e23

Commits on Aug 9, 2018

• Linux didn't like being flushed so much.

  dbolya

  committed
  5c39d8c

  Copy full SHA for 5c39d8c

Pagination

Previous

Next