Dropout: Use upcaling at training
cbovar opened this issue · comments
Cédric Bovar commented
Scale by 1 / (1 - dropProb) to have a magnitude constant accross training (dropProb > 0) and testing (dropProb = 0)
Deep Learning in C#
cbovar opened this issue · comments
Scale by 1 / (1 - dropProb) to have a magnitude constant accross training (dropProb > 0) and testing (dropProb = 0)