icevision

Supports pytorch 1.8

Added

• iou_thresholds parameter to COCOMetric
• SimpleConfusionMatrix Metric
• Negative samples support for yolov5
• Negative samples support for mmdetection object detection models

Changed

• Breaking: Albumentations aug_tfms defaults.
  • rotate_limit changed from 45 to 15
  • rgb_shift_limit changed from 20 to 10
  • VOC parser uses image sizes from annotation file instead of image
  • bumps fastai to latest version (<2.4)

[0.7.0]

BREAKING: API Refactor

Added

• Metrics for mmdetection models

Changed

• Breaking: Renamed tasks default,detect,classif to common,detection,classification
• Breaking: Renamed imageid to record_id
• Breaking: Added parameter is_new to Parser.parse_fields
• Removed all dependencies on cv2 for visualisation
• Use new composite API for visualisation - covers user defined task names & multiple tasks
• Added a ton of visualisation goodies to icevision.visualize.draw_data.draw_sample - user can now
  • use custom fonts
  • control mask thickness
  • control mask blending
  • prettify labels -- show confidence score & capitalise label
  • plot specific and/or exclude specific labels
  • pass in a dictionary mapping labels to specific colors
  • control label height & width padding from bbox edge
  • add border around label for legibility (color is a parameter)

Breaking:: Rename labels->label_ids, labels_names->labels in LabelsRecordComponent

• Renamed torchvision resnet backbones:
  • resnet_fpn.resnet18 -> resnet18_fpn
  • resnest_fpn.resnest18 -> resnest18_fpn

Breaking: Added parameters sample and keep_image to convert_raw_prediction
Breaking: Renamed VocXmlParser to VOCBBoxParser and VocMaskParser to VOCMaskParser
Breaking: Renamed predict_dl to predict_from_dl

[0.5.2]

Added

• aggregate_records_objects function

Changed

• Added label_field to VIA parser to allow for alternate region_attribute names

[0.5.0]

Added

• Keypoints full support: data API, model and training
• VGG Image Annotator v2 JSON format parser for bboxes
• figsize parameter to show_record and show_sample

Changed

• improved visualisation for small bboxes
• COCOMetric now returns all metrics from pycocotools
• makes torchvision models torchscriptable

[0.4.0]

Added

• retinanet: model, dataloaders, predict, ...

Changed

• Breaking: models/rcnn renamed to models/torchvision_models
• tests/models/rcnn renamed to tests/models/torchvision_models

Added

• pytorch 1.7 support, all dependencies updated
• tutorial with hard negative samples
• ability to skip record while parsing

Changed

• show_preds visual improvement