Hi,
I am reading caffe2 c++ codes: operators/rnn/recurrent_network_op.h.
In RecurrentNetworkOp::DoRunWIthType, the Op created a collection of workspaces, one for each t. Why each time step needs unique workspace? Seemed the hidden_state and cell_state and other recurrent states (both forward pass and backward pass) have been assigned enough storage for states of all the time steps?
Thank you very much!