jeffdonahue / lrcn Goto Github PK

Hello,

I thought while training you have to feed data and labels to the net so the solver can teach the net how input is connected to the output, but that isnt the way it works... in fact the input data is completly ignored while training and the sequence is considered to only depend on the timestep and previous labels. After i investigated this a little closer i found out that in no example the input data is fed to the net. https://github.com/junhyukoh/caffe-lstm/blob/master/examples/lstm_sequence/lstm_sequence.cpp

This seems to me as this is not the intended way how it should work.