search

torch / nngraph

Graph Computation for nn
Other
299 stars 96 forks source link

How to copy a node from a gmodule to another gmodule? #142

Open royal-feng opened 7 years ago

royal-feng commented 7 years ago

I want to do something like copy a node from a gmodule called net_A to another gmodule called net_B,so I do such operations netB.modules[i] = netA.modules[counter] and when the forward in netB, it failed.And the error imformation like that: torch/install/bin/luajit: ...torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:32: Only Cuda supported duh! stack traceback: [C]: in function 'assert' ...torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:32: in function 'resetWeightDescriptors' ...torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:96: in function 'checkInputChanged' ...torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:120: in function 'createIODescriptors' ...torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:188: in function 'func' /home/zsf/torch/install/share/lua/5.1/nngraph/gmodule.lua:345: in function 'neteval' /home/zsf/torch/install/share/lua/5.1/nngraph/gmodule.lua:380: in function 'forward' test3.lua:335: in main chunk [C]: in function 'dofile' .../zsf/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk [C]: at 0x00406670

So I try to copy the node.weight from netA to netB,like that netB.modules[i].weight=netA.modules[counter].weight netB.modules[i].bias=netA.modules[counter].bias But the result after netB:forward(input) is not so good. Is my copy operation something wrong? Thank you for answer!

arthitag commented 7 years ago

Hi, I don't have an answer to your question but I am facing a similar issue. I am trying to copy weights from a trained nngraph model to a new nngraph model (with almost same structure) that I created from file. When I do a forward pass results are all wrong and I can't figure out why. For copying weights I used params0, gp0 = model0:parameters() params1, gp1= model1:parameters() and then serially copied all weight/bias tensors in wt to the params tensors on model 1 using the tensor1:copy(tensor2). I later verified if the weights of two models are the same, and if the inputs are getting passed correctly. I don't know why but some of the intermediate layer activations in the forward pass are growing really large and then going to all 0's.