About the CE loss and the preprocess. #17
Description
Hi, recently,i get some trouble about how to use the Crossentropyloss,i have preprocessed the brats2017 dataset according your preprocess.py when i train the model, i find the label's shape is [bs,3,128,128,128] and its type is float32, so i read your preprocess file in the file the label is get through the code '' result = [(img == 2) | (img == 3), (img == 2) | (img == 3) | (img == 1), img == 3]'',but the CEloss target is limited one channel .such as ce(out,label)the out.shape is [4,3,128,128,128]and the label must be [4,128,128,128].but in this label,i can't use argmax to transfer the shape.so can you help me to solve the problem? thanks!