Added

• Overview document for feature highlights in v0.6
• 10 new transforms, a masked loss wrapper, and a NetAdapter for transfer learning
• APIs to load networks and pre-trained weights from Clara Train Medical Model ARchives (MMARs)
• Base metric and cumulative metric APIs, 4 new regression metrics
• Initial CSV dataset support
• Decollating mini-batch as the default first postprocessing step
• Initial backward compatibility support via monai.utils.deprecated
• Attention-based vision modules and UNETR for segmentation
• Generic module loaders and Gaussian mixture models using the PyTorch JIT compilation
• Inverse of image patch sampling transforms
• Network block utilities get_[norm, act, dropout, pool]_layer
• unpack_items mode for apply_transform and Compose
• New event INNER_ITERATION_STARTED in the deepgrow interactive workflow
• set_data API for cache-based datasets to dynamically update the dataset content
• Fully compatible with PyTorch 1.9
• --disttests and --min options for runtests.sh
• Initial support of pre-merge tests with Nvidia Blossom system

Changed

• Base Docker image upgraded to nvcr.io/nvidia/pytorch:21.06-py3 from nvcr.io/nvidia/pytorch:21.04-py3
• Optionally depend on PyTorch-Ignite v0.4.5 instead of v0.4.4
• Unified the demo, tutorial, testing data to the project shared drive, and Project-MONAI/MONAI-extra-test-data
• Unified the terms: post_transform is renamed to postprocessing, pre_transform is renamed to preprocessing
• Unified the postprocessing transforms and event handlers to accept the "channel-first" data format
• evenly_divisible_all_gather and string_list_all_gather moved to monai.utils.dist

Removed

• Support of 'batched' input for postprocessing transforms and event handlers
• TorchVisionFullyConvModel
• set_visible_devices utility function
• SegmentationSaver and TransformsInverter handlers

Fixed

• Issue of handling big-endian image headers
• Multi-thread issue for non-random transforms in the cache-based datasets
• Persistent dataset issue when multiple processes sharing a non-exist cache location
• Typing issue with Numpy 1.21.0
• Loading checkpoint with both model and optmizier using CheckpointLoader when strict_shape=False
• SplitChannel has different behaviour depending on numpy/torch inputs
• Transform pickling issue caused by the Lambda functions
• Issue of filtering by name in generate_param_groups
• Inconsistencies in the return value types of class_activation_maps
• Various docstring typos
• Various usability enhancements in monai.transforms