Commits

Breadcrumbs

History for

ononnx

User selector

All users

Datepicker

All time

Commit History

Commits on Sep 22, 2018

• Added an eval flag (--mask_proto_debug) to output the files needed for compute_masks and made the script more streamlined. It's sorta hacky though, but you've gotta do what you've gotta do.

  dbolya

  committed
  0fc359b

  Copy full SHA for 0fc359b

Commits on Sep 21, 2018

• Fixed the overexposure of the mask drawn with cv2. No wonder it was over exposed, I forgot to convert the colors from [0, 255] to [0, 1].

  dbolya

  committed
  7fab5ca

  Copy full SHA for 7fab5ca

• Added the ability to seed the eval rng.

  dbolya

  committed
  1c9291a

  Copy full SHA for 1c9291a

Commits on Sep 20, 2018

• Added a --no_sort argument to eval that doesn't do the whole sorting by hashed image id I added earlier. This is so we can produce web detections with the same images as before.

  dbolya

  committed
  864bc59

  Copy full SHA for 864bc59

Commits on Sep 14, 2018

• 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 13, 2018

• Do a deterministic shuffle when selecting the top n validation images even if shuffle is off. I do this because the order of dataset.ids is not deterministic and differs between python version. Wha…

  Show description for fe3ee4e
  dbolya

  committed
  fe3ee4e

  Copy full SHA for fe3ee4e

• Made some eval code more legible and fixed display spamming the console.

  dbolya

  committed
  88e8c5b

  Copy full SHA for 88e8c5b

Commits on Sep 12, 2018

• Created a script to mix different datasets together without having to copy images. Also added config settings and readme instructions for trainval35k and minival5k.

  dbolya

  committed
  eda0294

  Copy full SHA for eda0294

• Finally cleaned up that function like I said I would.

  dbolya

  committed
  821adc8

  Copy full SHA for 821adc8

Commits on Sep 11, 2018

• Added a benchmark mode to the eval script.

  dbolya

  committed
  3396424

  Copy full SHA for 3396424

Commits on Sep 8, 2018

• Turns out timing information wasn't including a big synch point. I've fixed that and added a catch-all timing environment, but that means are network was actually 15 fps this whole time.

  dbolya

  committed
  b732390

  Copy full SHA for b732390

Commits on Aug 31, 2018

• Added a transform config that allows us to properly preprocess the image for VGG and Resnet differently, fixing a bug where resnet was getting vgg-processed

  dbolya

  committed
  beb805d

  Copy full SHA for beb805d

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

• Fixed a bug where train.py was using cross-class nms for validation.

  dbolya

  committed
  bf4a3ec

  Copy full SHA for bf4a3ec

Commits on Aug 25, 2018

• Added the second round of ablations and implemented all their required features. The jist of it is that I added a standardized way to modify the 3 nonlinearities and added prototype loss.

  dbolya

  committed
  a4f5a91

  Copy full SHA for a4f5a91

• Finally added a shortcut to parse the config name from the training file name if --config is left blank.

  dbolya

  committed
  eb9be3c

  Copy full SHA for eb9be3c

• Added an eval flag to visualize the way a mask is put together when using lincomb.

  dbolya

  committed
  6ffd98b

  Copy full SHA for 6ffd98b

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 15, 2018

• Added the capability for deconv layers in the proto branch and added some configs.

  dbolya

  committed
  ff03798

  Copy full SHA for ff03798

Commits on Aug 14, 2018

• Sped up some eval-time stuff by making the cpu and gpu stuff more consistent.

  dbolya

  committed
  d1cac8b

  Copy full SHA for d1cac8b

Commits on Aug 11, 2018

• Added support for multiple mask output types, added lincomb support, moved postprocessing out of eval into one massive nice new function (which still needs to be optimized), and added some configs …

  Show description for e55051f
  dbolya

  committed
  e55051f

  Copy full SHA for e55051f

Commits on Aug 10, 2018

• Changed ssd550_resnet101_yolo_matching_config to not train masks.

  dbolya

  committed
  3e4de17

  Copy full SHA for 3e4de17

Commits on Aug 9, 2018

• Added a web-based detections viewer and the capability for eval to produce detections for it.

  dbolya

  committed
  f07e7ee

  Copy full SHA for f07e7ee

Commits on Aug 7, 2018

• Fixed up the bbox scaling on the eval script and cleaned some code.

  dbolya

  committed
  a045267

  Copy full SHA for a045267

Commits on Aug 6, 2018

• Added the ability to select the config to use in the eval script.

  dbolya

  committed
  45a73c2

  Copy full SHA for 45a73c2

Commits on Aug 2, 2018

• Added the ability to change configs as a command-line argument, and added a config for ssd600_resnet101 which is just as it sounds.

  dbolya

  committed
  6637e31

  Copy full SHA for 6637e31

• If preserve_aspect_ratio is set, during eval with --display, the images are now cropped properly and detections outside of the crop are ignored.

  dbolya

  committed
  37105a0

  Copy full SHA for 37105a0

Commits on Jul 27, 2018

• Added a SavePath class which handles parsing and generating save paths for the model. I've also made it so you can do --resume=interrupt and --trained_model=interrupt to load the interrupt model au…

  Show description for 581c791
  dbolya

  committed
  581c791

  Copy full SHA for 581c791

• Added a 'use_prediction_module' flag in the config file to enable / disable the prediction module (c) from DSSD.

  dbolya

  committed
  62d800e

  Copy full SHA for 62d800e

Commits on Jul 26, 2018

• Added the ability to change out the backbone network and made it easily configurable. Also implemented a VGG backbone and added some settings to the training script. Note: in progress.

  dbolya

  committed
  87eda7e

  Copy full SHA for 87eda7e

Commits on Jul 24, 2018

• Created a script that can optimize the average IoU of anchor boxes instead of relying on kmeans to hopefully do that for us.

  dbolya

  committed
  67b25e3

  Copy full SHA for 67b25e3

• Fixed a bunch of bugs:

  Show description for 24a4686
  dbolya

  committed
  24a4686

  Copy full SHA for 24a4686

Commits on Jul 21, 2018

• Completely changed everything about the project (ie implemented the things in the readme). I have no idea if any of this works or why the maximum batch size I can fit in memory is 4. This is still …

  Show description for 7b2e7e4
  dbolya

  committed
  7b2e7e4

  Copy full SHA for 7b2e7e4

Commits on Jul 18, 2018

• Renamed eval_coco.py to run_coco_eval so I could tab complete eval.py. Also changed the threshold for masks from np.average(mask) to 0.5 because I am using sigmoid here after all.

  dbolya

  committed
  dfeb80c

  Copy full SHA for dfeb80c

Pagination

Previous

Next