HISTORY.md

History

• master
• 0.9.0 (2019-07-30)
  • make image transforms part of preprocessing #100
  • field-based two-scale implementation #101, also modifies single-scale decoder
  • added a show.CrowdPainter to visualize crowd annotations
  • new benchmark script and updated performance numbers in README.md, #104
• 0.8.0 (2019-07-08)
  • add support for resnext50, shufflenetv2x1 and shufflenetv2x2
  • new pretrained models
  • new transforms.RandomApply() and transforms.RotateBy90(); removed old transforms
  • new blur augmentation
  • improved BCE masks #87
• 0.7.0 (2019-06-06)
  • faster seed generation in decoder
  • training log plot improvements (labels, consistent colors)
  • improved debug visualizer for decoder
• 0.6.3 (2019-05-28)
  • support parallel decoding for predict and eval_coco (~4x speed improvement) which is automatically activated for batch sizes larger than 1
• 0.6.2 (2019-05-23)
  • improved decoder performance #61, #63, #64
  • remove apply_class_sigmoid property from head nets and use the standard model.train() and model.eval() methods instead
  • bugfix for runs with padding
  • improved log messages
  • log the names of the fields (preparation to improve plots of training losses)
• 0.6.1 (2019-05-13)
  • improved decoder performance #51
  • experiments with ONNX
  • MultiHeadLoss #50 (no external API changes)
  • automatically add hash to trained model files for Model Zoo compatibility #53
  • support nested objects in factories #56
• 0.6.0 (2019-05-07)
  • Torch 1.1.0 compatibility (older versions work but have modified learning rate schedule due to pytorch/pytorch#7889)
  • more aggressive NMS
  • multi-scale support for eval_coco: --two-scale, --three-scale, --multi-scale
  • guaranteed complete poses with --force-complete-pose
• 0.5.1 (2019-05-01)