v0.5: add distribution information (#62)

• Add distribution information of the package
• Add support for regular expressions in pocket.utils.ImageHTMLTable
• Update installation instructions

Add support for automatic mixed precisions in learning engines

The preferable PyTorch version is now 1.9.0.

Update log:

• Add support for automatic mixed precisions
• Multiple bug fixes

Some new features and more bug fix

This will be the last release based on PyTorch 1.1.0.

Update log:

• Implement new class pocket.data.DatasetConcat to allow joining multiple datasets
• Allow single box (pair) in pocket.utils.draw_boxes and pocket.utils.draw_box_pairs
• Fix an issue related to multiprocessing (#32)
• Fix incorrectly implemented repr methods
• Other insignificant things

New release with more features and bug fix

• New classes pocket.utils.BoxAssociation and pocket.utils.BoxPairAssociation that handles association for detection tasks
• Visualisation of bounding box pairs with pocket.utils.draw_box_pairs
• Fixed a bug in pocket.utils.AveragePrecisionMeters and pocket.utils.DetectionAPMeter regarding handling division by zero
• Allow the precision for float-point data in average precision meters
• Allow single process when computing average precision to avoid an error related to the number of file descriptors
• Remove obsolete object classes in MS COCO (91->80) for Faster R-CNN
• New class pocket.ops.RandomHorizontalFlip that handles both images and bounding boxes
• Include learning rate scheduler as a state variable in learning engines
• A few other minor changes

Initial release

Initial release of Pocket, featuring versatile learning engines