ONNX Export, Webcam Inference, MultiScale Training

This release requires PyTorch >= v1.0.0 to function properly. Please install the latest version from https://github.com/pytorch/pytorch/releases

Breaking Changes

There are no breaking changes in this release.

Bug Fixes

• N/A

Added Functionality

• MultiScale Training #52: train.py --multi-scale will train each batch at a randomly selected image size from 320 to 608 pixels.
• Webcam Inference #89: set webcam=True in detect.py.
• Video Inference. Pass a video file to detect.py.
• YOLOv3-tiny support #51: detect.py --cfg cfg/yolov3-tiny.cfg --weights weights/yolov3-tiny.pt
• YOLOv3-spp support #16. detect.py --cfg cfg/yolov3-spp.cfg --weights weights/yolov3-spp.pt
• ONNX Export #82: ONNX_EXPORT = True in models.py.

Performance

• 20% improvement in training speed via code optimization and removal of redundant batch_report functionality. All of this functionality, including computation of TP, FP, FN, Precision, Recall and mAP is now done in test.py after each training epoch.

TODO (help and PR's welcome!)

• Multi GPU support #21.
• YAPF linting (including possible wrap to PEP8 79 character-line standard) #88.
• test.py should ideally output text files in the official COCO mAP format as well for external mAP computation #2 (comment).

Initial Release

This is an initial release. This repository currently works well for inference, with no known issues, with training still under development. Current COCO training mAP using this repo is 0.522 (at 416 x 416) after 62 epochs (using all default training settings, simply running python3 train.py). We are exploring ways to improve this further.

Loss curves, Precision, Recall and mAP.

Recommend PyTorch >0 v1.0.0 to run this repo, which includes numerous bug fixes: https://github.com/pytorch/pytorch/releases

Inference

• Inference appears to be working well, no known issues.

Training

• Training is still under development. v1.0 benchmark is currently 0.522 mAP at epoch 62 at 416 image size without multi-scale training.
• Balancing of the various loss terms seem to have great effect on the results, the current constants were tuned for training performance. Further study is needed here.
• Augmentation may be a bit aggressive, experimentation is needed at reduced levels.
• Training seems to suffer jumps in losses (visible in plots above). This may be associated with restarting a stopped training session, the cause is unclear, and some users seem not to experience this.

Validation

• mAP calculation is now correct. mAP is calculated per class per image, and then averaged over images. Despite this, current mAP calculation in test.py seems to be slightly different than the official COCO mAP code.
• It would be very useful for test.py to additionally output text files in the official COCO mAP format.

Multi-GPU

• Multi GPU is currently not supported. PRs welcome!

Performance

• Numerous commits addressed performance issues. Largest changes produced by not reporting P, R, TP, FP, FN each batch, and by reducing data movement between GPU and CPU. Current training speed is about 1 epoch per hour (16-image batches at 416 x 416 each take about 0.7 seconds) on a GCP P100.