This document will detail all changes I make. I don't know how I'm going to be versioning things yet, so you get dates for now.
2020.01.25:
- Fixed the mask IoU branch crashing when all masks in a batch are discarded (fixes #302, #259).
2020.01.24:
- Fixed the conv layer detection during initialization to work with pytorch 1.4 (fixes #292).
2020.01.23:
- Fixed the video playback crashing if there's nothing in the scene (fixes #266).
- Fixed the logger logging the last loss as total loss instead of the actual total (fixes #254).
2019.12.16 (v1.2):
- Added YOLACT++ implementation, paper, and code.
- Added DCN support (need to compile CUDA kernels if you want to use them, see README).
- Added a mask rescoring network trained with mask iou.
- Added configs with more anchors.
2019.12.06:
- Made training much more stable (no more infs and hopefully fewer loss explosions) by ignoring
augmented boxes with < 4px of height and width (this includes 0 area boxes which caused the inf).
See #222 for details.
2019.11.20:
- Fixed bug where saving videos wouldn't work when using cv2 not compiled with display support (#197).
2019.11.06:
- Changed Cython import to only active when using traditional nms.
- Added cross-class fast NMS.
2019.11.04:
- Fixed a bug where the learning rate auto-scaling wasn't being applied properly.
- Fixed a logging bug were lr was sometimes not properly logged after a resume.
2019.10.25 (v1.1):
- Added proper Multi-GPU support. Simply increase your batch size to 8*num_gpus and everything will scale.
- I get an ~1.8x training speed increase when using 2 gpus and an ~3x increase when using 4.
- Added a logger that logs everything about your training.
- Check the Logging section of the README to see how to visualize your logs. (Not written yet)
- Savevideo now uses the evalvideo framework and suports --video_multiframe. It's much faster now!
- Added the ability to display fps right on the videos themselves by using --display_fps
- Evalvideo now doesn't crash when it runs out of frames.
- Pascal SBD is now officially supported! Check the training section for more details.
- Preserve_aspect_ratio kinda sorta works now, but it's iffy and the way I have it set up doesn't perform better.
- Added a ton of new config settings, most of which don't improve performance :/
2019.09.20
- Fixed a bug where custom label maps weren't being applied properly because of global default argument initialization.
2019.08.29
- Fixed a bug where the fpn conv layers weren't getting initialized with xavier since they were being overwritten by jit modules (see #127).
2019.08.04
- Improved the matching algorithm used to match anchors to gt by making it less greedy (see #104).
2019.06.27
- Sped up save video by ~8 ms per frame because I forgot to apply a speed fix I applied to the other modes.