CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog.

Unreleased - 0.10.0a1

Main dependencies updated

• torch 1.9.0
• tochvision 0.10
• mmdet 2.16.0
• mmcv 1.3.14
• fastai 2.5.2
• pytorch-lightning 1.4.8

Unreleased - 0.9.0a1

Added

• Low level parsing workflow with RecordCollection
• Semantic segmentation support with fastai

Changed

• Breaking: Refactored mask components workflow
• Breaking: Due to the new mask components refactor, autofix doesn't work for mask components anymore.

[0.8.1]

Added

• end2end_detect(): Run Object Detection inference (only bboxes) on a single image, and return predicted boxes corresponding to original image size -Breaking: BaseLabelsRecordComponent as_dict() now returns both labels and labels_ids. labels are now strings instead of integers.

Changed

• Breaking: On tfms.A.aug_tfms parameter size and presize changed from order (height, width) to (width, height)
• Added RecordCollection
• Breaking: Changed how the resnet (not-fpn) backbone cut is done for torchvision models. Previous resnet torchvision trained models will have trouble loading weights.

[0.8.0]

Supports pytorch 1.8

Added

• iou_thresholds parameter to COCOMetric
• SimpleConfusionMatrix Metric
• 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

[0.6.0b1]

Added

• mmdetection models

Changed

• Breaking: All Parser subclasses need to call super.__init__
• Breaking: LabelsMixin.labels now needs to return List[Hashable] instead of List[int] (labels names instead of label ids)
• Breaking: Model namespace changes e.g. faster_rcnn -> models.torchvision.faster_rcnn, efficientdet -> models.ross.efficientdet
• Breaking: Renamed ClassMap.get_name/get_id to ClassMap.get_by_name/get_by_id
• Breaking: Removes idmap argument from Parser.parse. Instead pass idmap to the constructor (__init__).
• ClassMap is not created inside of the parser, it's not required to instantiate it before
• class_map labels get automatically filled while parsing
• background for class_map is now always 0 (unless no background)
• adds class_map to Record

Deleted

[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

[0.3.0]

Added

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

Changed

• show_preds visual improvement

[0.2.2]

Added

• Cache records after parsing with the new parameter cache_filepath added to Parser.parse (#504)
• Added pretrained: bool = True argument to both faster_rcnn and mask_rcnn model() methods. (#516)
• new class EncodedRLEs
• all masks get converted to EncodedRLEs at parsing time

Changed

• Removed warning on autofixing masks
• RLE default counts is now COCO style
• renamed Mask.to_erle to Mask.to_erles

[0.2.1]

Changed

• updated matplotlib and ipykernel minimum version for colab compatibility

[0.2.0]

IMPORTANT

• Switched from poetry to setuptools

Added

• Function wandb_img_preds to help logging bboxes to wandb
• wandb as a soft dependency
• Template code for parsers.SizeMixin if parsers.FilepathMixin is used
• Get image size without opening image with get_image_size
• Ability to skip record while parsing with AbortParseRecord
• Autofix for record: autofix_records function and autofix:bool parameter added to Parser.parse
• Record class and mixins, create_mixed_record function to help creating Records
• InvalidDataError for BBox
• Catches InvalidDataError while parsing data

Changed

• Breaking: Unified parsers.SizeMixin functions image_width and image_height into a single function image_width_height
• Rename Parser SizeMixin fields from width height to image_width image_height

Deleted

• Removed CombinedParser, all parsing can be done with the standard Parser

[0.1.6]

Added

• Efficientdet now support empty annotations

Changed

• Returns float instead of dict on FastaiMetricAdapter

[0.1.5]

Changed

• Updates fastai2 to the final release version

[0.1.4]

Added

• soft import icedata in icevision.all
• show_pbar parameter to COCOMetric

Changed

Deleted

[0.1.3]

Changed

• Effdet as direct dependency

[0.1.2]

Added

• show_results function for each model

Changed

• Default data_splitter for Parser changed to RandomSplitter
• Renamed package from mantisshrimp to icevision

Deleted

• Removed datasets module to instead use the new icedata package

[0.0.9]

Added

• batch, samples = <model_name>.build_infer_batch(dataset)
• preds = <model_name>.predict(model, batch)
• infer_dl = <model_name>.infer_dataloader(dataset)
• samples, preds = predict_dl(model, infer_dl)
• Dataset.from_images Contructs a Dataset from a list of images (numpy arrays)
• tfms.A.aug_tfms for easy access to common augmentation transforms with albumentations
• tfms.A.resize_and_pad, useful as a validation transform
• **predict_kwargs to predict_dl signature
• from mantisshrimp.all import * to import internal modules and external imports
• show parameter to show_img
• download_gdrive and download_and_extract_gdrive
• New datasets pennfundan and birds

Changed

• Renames AlbuTransform to AlbumentationTransforms
• All build_batch method now returns batch, samples, the batch is always a tuple of inputs to the model
• batch_tfms moved to tfms.batch
• AlbumentationTransforms moved to tfms.A.Adapter
• All parsers function were moved to their own namespace parsers instead of being on the global namespace so, for example, instead of Parser now we have to do parsers.Parser
• Removed Parser word from Mixins, e.g. ImageidParserMixin -> parsers.ImageidMixin
• Removed Parser word from parser default bundle, e.g. FasterRCNNParser -> parsers.FasterRCNN
• COCO and VOC parsers moved from datasets to parsers
• DataSplitters moved from parsers/splits.py to utils/data_splitter.py
• Renames *_dataloader to *_dl, e.g. mask_rcnn.train_dataloader to mask_rcnn.train_dl
• Moves RecordType from parsers to core
• Refactors IDMap, adds methods get_name and get_id
• Moves IDMap from utils to data
• DataSplitter.split now receives idmap instead of ids

[0.0.0-pre-release]

Added

• CaptureStdout for capturing writes to stdout (print), e.g. from COCOMetric
• mantisshrimp.models.<model_name>.convert_raw_predictions to convert raw preds (tensors output from the model) to library standard dict
• COCOMetricType for selecting what metric type to use (bbox, mask, keypoints)
• COCOMetric fixed
• sort parameter for get_image_files
• ClassMap: A class that handles the mapping between ids and names, with the optional insertion of the background class

Changed

• All dataloaders now return the batch and the records, e.g. return (images, targets), records
• Metric.accumulate signature changed to (records, preds), reflects in FastaiMetricAdapter and LightningModelAdapter
• datasets.<name>.CLASSES substituted by a function datasets.<name>.class_map that returns a ClassMap
• datasets.voc.VocXmlParser, show methods: parameter classes: Sequence[str] substituted by class_map: ClassMap
• datasets.fridge.parser, datasets.pets.parser: additional required parameter class_map

Removed

• MantisFasterRCNN, MantisMaskRCNN
• MantisEfficientDet
• CategoryMap, Category
• MantisModule

Links