After the PR torch/nn#526 , native
clearState()method does clear intermediate buffers, so you no longer need this snippet.
During forward/backward propagation, some of the nn/cunn modules utilize temporary buffers. Those buffers have no meaninful information after training and saving such model wastes too much of disk space. This script can be used to free unnecessary buffers in the model before you save it onto disk.
Refer to or run the script example.lua.