Any neat way to debug?

Question

Any neat way to debug?

blackyang opened this issue 8 years ago · comments

Is there any neat way to debug nngraph? Thanks!

Currently I write a Debug layer which is almost the same as Identity layer except that I can print whatever I want inside updateOutput() and updateGradInput()

Xiao Yang · Answer 1 · Sun Mar 27 2016 06:36:19 GMT+0800 (China Standard Time)

seems that there is a similar implementation in dpnn, the PrintSize layer:)

Joost van Doorn · Answer 2 · Fri Apr 08 2016 05:12:37 GMT+0800 (China Standard Time)

You can also add more outputs. Like:

input = nn.Identity()()
L1 = nn.Tanh()(nn.Linear(10, 20)(input))
L2 = nn.Tanh()(nn.Linear(30, 60)(nn.JoinTable(1)({input, L1})))
L3 = nn.Tanh()(nn.Linear(80, 160)(nn.JoinTable(1)({L1, L2})))

g = nn.gModule({input}, {L1,L2,L3})

Xiao Yang · Answer 3 · Fri Apr 08 2016 05:15:49 GMT+0800 (China Standard Time)

@JoostvDoorn Thanks! But in that way we need to provide some zeroTensors as gradOutput when during backprop

Xiao Yang · Answer 4 · Sat Sep 23 2017 00:27:32 GMT+0800 (China Standard Time)

FYI, pytorch introduces similar functionalities: register_forward_hook and register_backward_hook