Skip to content

Latest commit

 

History

History
73 lines (72 loc) · 5.04 KB

HISTORY.md

File metadata and controls

73 lines (72 loc) · 5.04 KB
Raw

History

  • master
  • 0.11.6 (2020-06-15)
    • circular truncation for CifHr
    • extend Guide
  • 0.11.4 (2020-06-02)
    • init Guide
    • minor updates
  • 0.11.3 (2020-06-01)
    • new debug message for "neural network device" (to check cpu versus gpu usage)
    • debug output without plots is default; enable debug plots with new --debug-images
  • 0.11.2 (2020-05-29)
    • pretrained resnet50 model
    • fix CUDA support in openpifpaf.video
    • add --version option to all CLIs
  • 0.11.0 (2020-05-12)
    • major refactor
    • now requires Python>=3.6 for type annotations
    • new ShuffleNetV2 models: shufflenetv2k16w and shufflenetv2k30w
    • 64bit loss and Focal Loss for confidences
    • fast fused convolutions for CompositeHeadFused
    • new handling of crowd annotations in encoder
    • new --extended-scale training and eval-coco option
    • decoding with frontier is default
    • more robust blending of connection candidates
    • introduced openpifpaf.visualizer and many improvements to visualization
    • [experimental] new cocodet dataset interface for detections
  • 0.10.1 (2019-12-09)
    • faster decoder
    • refactored scale generation between loss and encoder
  • 0.10.0 (2019-10-21)
    • major refactor: move all factory-code into factories
    • new experimental decoder
    • improved image rescaling randomization
    • module-level logging
    • index-matching bug fixed by @junedgar #147
    • tests for Windows, PyTorch 1.3 and with pylint 2.4
  • 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)