The split used by Pylearn2 for training, validation and test in dense_dataset relies on an invariant order of training image paths from the settings file. Currently these are just sorted to make sure they're always the same. It would be good to use a random seed and then shuffle at this point, from the point of view of the Pylearn2 dataset split. Having such a deterministic split may be causing us problems.
The split used by Pylearn2 for training, validation and test in dense_dataset relies on an invariant order of training image paths from the settings file. Currently these are just sorted to make sure they're always the same. It would be good to use a random seed and then shuffle at this point, from the point of view of the Pylearn2 dataset split. Having such a deterministic split may be causing us problems.