FixableDropout
PyTorch Implementation of Dropout layers with fixable masks
To allow the computation with a fixed computational graph in a neural network with dropout layers, we provide fixable Dropout in 1D/2D/3D: FixableDropout1d, FixableDropout2d and FixableDropout3d. One can simply replace nn.Dropout1d/nn.Dropout2d/nn.Dropout3d with them. Intead of directly saving last generated masks, we keep a record of the mask generator, which is parameter efficient.
contact:
Chen Chen, Imperial College London. chen.chen15@imperial.ac.uk
Please star our project if you like it and find it useful.