Lstm in c++/training tips

A developer reports issues with their LSTM implementation in C++, noting that training error initializes around 4.5 but can spike into the 20s, which they have not seen with other networks. They suspect a matrix indexing swap and indicate they may stop ML programming due to hardware limitations.

i think there’s an issue with my implementation since my experience is disparate with others. unfortunately i hid my second post by using an out of fashion expression. it might read better for people with wider monitors than i. or it’s in rnn-lstm.h at git hub. com/atomictraveller/scary-mess must be negative log likelihood as training often initialises around 4.5. “my lstm” can send error into the 20s and up which i don’t think i’ve seen before with any net. i have toxostomas that tweet when i think about math or sleep heheh. which explains why j i is swapped i think this is as far as i go with ML programming, i enjoy generation, my hardware in a number of senses is probably better scaled for procedural development.