Open NathanMolinier opened 1 year ago
Average is tricky (bc if binary hot-encoded labels that wouldn't be compatible)-- other stragegy: majority voting-- MONAI or pytorch has that-- see example https://github.com/ivadomed/model_seg_mouse-sc_wm-gm_t1/blob/f5750a9c6cd85c3b5d3d08a15b2ee3820c53440b/test.py#L98-L106
also read other implementations
Description
Currently, the hourglass network is trained on a single "fold" corresponding to only one split of a dataset (train/validation/test). To improve the ability of the network to generalise, multiple splits could be used for the same dataset and then averaged to improve the robustness of the predictions.