Problem with nn.Replicate(1) about cudnn.torch

[fmassa]

@xonobo can't you just take your input and do

input=input:view(1,input:size(1),input:size(2),input:size(3))

before feeding the input to the network ?

[mys007]

Some time ago I observed that cudnn doesn't like 0-strides, I think this is the case. fmassa's solution should work.

[xonobo]

I tried this but now gives an assertion fail

m:forward(torch.rand(1,7,7):view(1,1,7,7):cuda())
/torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:92: assertion failed!
stack traceback:
[C]: in function 'assert'
/torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:92: in function 'createIODescriptors'
/torch/install/share/lua/5.1/cudnn/SpatialConvolution.lua:339: in function 'updateOutput'
/torch/install/share/lua/5.1/nn/Sequential.lua:44: in function 'forward'
[string "_RESULT={m:forward(torch.rand(1,7,7):view(1,1..."]:1: in main chunk
[C]: in function 'xpcall'
/torch/install/share/lua/5.1/trepl/init.lua:650: in function 'repl'
/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:199: in main chunk

By the way I want to handle this dummy resizing in the model. Since the input sizes of the images may vary I could not use nn.View.

[fmassa]

@xonobo remove the nn.Replicate from your network.
Also, no need to use the nn.View module, apply the view directly to the input data, this way it's independent on the input size

[xonobo]

ok forgot to remove the replicate. Now it works.

Thanks for the solution. This is not what I exactly want but handle the resizing of input before feeding to the model.

[fmassa]

if you absolutely want to have the nn.Replicate(1) in your network, add also just after it a nn.Copy(nil,nil,true), and then you can feed the input without viewing it. But it won't handle the case where the input is already batched.