Unify the interface for datasets and models. Now one can call reid.models.create('resnet50', ...) and reid.datasets.create('cuhk03', ...) to create instances.
Reorganize examples by their methods rather than the net architecture, for example, softmax loss classification, triplet loss verification, and OIM loss. Make the dataset and net architecture command line parameters for the examples.
Now saving / loading checkpoints gets rid of DataParallel in examples.
Three main changes:
Unify the interface for datasets and models. Now one can call
reid.models.create('resnet50', ...)andreid.datasets.create('cuhk03', ...)to create instances.Reorganize examples by their methods rather than the net architecture, for example, softmax loss classification, triplet loss verification, and OIM loss. Make the dataset and net architecture command line parameters for the examples.
Now saving / loading checkpoints gets rid of DataParallel in examples.