the meaning of parameter inside lstm cell

[liao andrew]

hi all, 

i am now try to implement the latest cvpr work natural language object retrieval on tensorflow. however, i got stuck when using the pretrained model.

the code initialize-weights-scrc-full.py just use:

new_net.params['lstm2_context'][0].data[...] = old_net.params['lstm2'][0].data[...]

new_net.params['lstm2_context'][1].data[...] = old_net.params['lstm2'][1].data[...]

new_net.params['lstm2_context'][2].data[...] = old_net.params['lstm2'][2].data[...]

new_net.params['lstm2_context'][3].data[...] = old_net.params['lstm2'][3].data[...]

what i know is that it is trying to assign the old net parameters to the new net parameters. 

however, i know the what exactly the meaning(forget/input... gate matrix/bias?) of:

old_net.params['lstm2'][0].data[...] or old_net.params['lstm2'][1].data[...] ...

can any lstm master help me out?