v2.0.0-beta

This module now uses PyTorch and MMDetection instead of TensorFlow. The MAIA object detection is now implemented with Faster R-CNN instead of Mask R-CNN. In a few tests, this new setup performed a little better than before (higher recall and precision). The results were a lot more stable, too (i.e. recall and precision did not change that much when training/inference was repeated on the same data).

Upgrade information

The MAIA_AVAILABLE_BYTES env variable was removed. Instead, the object detection training batch size can be set directly with MAIA_MMDET_TRAIN_BATCH_SIZE.

The COCO_MODEL_URL env variable was removed. Use MAIA_BACKBONE_MODEL_URL for the PyTorch backbone file and MAIA_MODEL_URL for the MMDetection checkpoint file.

The training scheme is no longer configurable. Instead of iteration-based training (which still was called "epochs" before), training runs for a fixed number of 12 epochs now (which is the default training config for many MMdet object detectors and worked well in tests).

What's previously been called "instance segmentation" is now correctly called "object detection".

What's Changed

• Bump terser from 4.8.0 to 4.8.1 by @dependabot in #108
• Bump loader-utils from 1.4.0 to 1.4.1 by @dependabot in #113
• Bump loader-utils from 1.4.1 to 1.4.2 by @dependabot in #114
• Bump json5 from 1.0.1 to 1.0.2 by @dependabot in #116
• Migrate to PyTorch by @mzur in #117

Full Changelog: v1.24.0...v2.0.0-beta