Official version of smallNORB, with stereo images of size 96x96 Since the dataset is rather large 800+MB, the NORB_small class will dynamically load the train/valid or test datasets by overriding the __getattribute__ function. Any access to "train"/"valid" members, will: * if "train"/"valid" was not the member which was last accessed... * load the train data file from memory * resplit into the same train/valid splits as before * return reference to train or validation dataset Any access to the "test" member, will: * if "test" was not the member which was last accessed... * load the test data file from memory * return reference to test dataset Two consecutive access to train/valid or test will therefore have no access-time penalty. Alternating between .train and .test will be very costly however. Before loading the new dataset, the object tries to clean up all of the local references to the dataset. This is done in the hope that loading the test dataset will "replace" in memory the "train" dataset (and vice-versa). For this to be true however, the user should avoid having duplicate references to "train", "valid" or "test" members (or manually delete them before switching datasets) TODO: make dynamic loading optional

5f9ffefa9ca8040e18e1a69fdbbe34b8a19099bc sequencelabelling 20090130-rerun