ClTensor that cannot be copied back to CPU?

[szagoruyko]

I wanted to run a simple example with MNIST classification from here https://github.com/szagoruyko/loadcaffe/blob/master/examples/mnist_lenet.lua with cltorch, it runs fine after a few hacks but I cannot copy the output back to FloatTensor:

th> print{net.output}
{
  1 : ClTensor - size: 1x10
}
                                                                      [0.0001s]
th> net.output
/usr/local/bin/luajit: /usr/local/share/lua/5.1/torch/Tensor.lua:200: bad argument #1 to 'copy' (torch.*Tensor expected, got userdata)
stack traceback:
    [C]: in function 'copy'
    /usr/local/share/lua/5.1/torch/Tensor.lua:200: in function </usr/local/share/lua/5.1/torch/Tensor.lua:194>
    [C]: in function 'tostring'
    /usr/local/share/lua/5.1/trepl/init.lua:236: in function 'rawprint'
    /usr/local/share/lua/5.1/trepl/init.lua:268: in function 'print'
    /usr/local/share/lua/5.1/trepl/init.lua:632: in function 'repl'
    /usr/local/lib/luarocks/rocks/trepl/scm-1/bin/th:185: in main chunk
    [C]: at 0x010b389400

[hughperkins]

I believe you're loading cutorch after loading cltorch? Can you try loading cutorch first, and then loading cltorch? (cutorch monkey-patches the copy functions, and doesnt realize cltorch might have monkey-patched them too) (edit: that 'userdata' is a ClTensor, which is unrecognized by CudaTensor copy function)

[szagoruyko]

aah missed that, it works now, MNIST classification on HD Graphics:

registering spatialconvolutionmm
Successfully loaded /opt/caffe/examples/mnist/lenet_iter_10000.caffemodel
module 'mnist' not found
conv1: 20 1 5 5
conv2: 50 20 5 5
ip1: 1 1 800 500
ip2: 1 1 500 10
Using Apple platform: Apple
Using device: HD Graphics 4000
nn.Sequential {
  [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> output]
  (1): nn.SpatialConvolutionMM(1 -> 20, 5x5)
  (2): nn.SpatialMaxPooling(2,2,2,2)
  (3): nn.SpatialConvolutionMM(20 -> 50, 5x5)
  (4): nn.SpatialMaxPooling(2,2,2,2)
  (5): nn.View
  (6): nn.Linear(800 -> 500)
  (7): nn.ReLU
  (8): nn.Linear(500 -> 10)
}
statefultimer v0.6
ConfusionMatrix:
[[     973       0       1       0       0       0       1       2       3       0]   99.286%
 [       0    1130       2       1       0       0       1       1       0       0]   99.559%
 [       1       2    1026       0       0       0       0       2       1       0]   99.419%
 [       0       0       1    1003       0       3       0       0       3       0]   99.307%
 [       0       0       0       0     978       0       0       1       0       3]   99.593%
 [       2       0       0       8       0     879       1       0       2       0]   98.543%
 [       3       2       0       1       1       3     946       0       2       0]   98.747%
 [       0       2       6       1       0       0       0    1018       0       1]   99.027%
 [       1       0       2       1       0       1       1       0     967       1]   99.281%
 [       1       2       0       4       6       4       1       4       1     986]]  97.721%
 + average row correct: 99.048245549202%
 + average rowUcol correct (VOC measure): 98.12281191349%
 + global correct: 99.06%

[hughperkins]

Cool :-) By the way, we could modify the cutorch monkey-patching to allow loading in any order. But I guess I dont want to touch cutorch too much for now, in case I break something. You can see the monkey-patcher for cltorch here though: https://github.com/hughperkins/cltorch/blob/master/init.lua#L3-L34 Basically it saves the old functions, and calls those, unless it sees an incoming ClTensor, in which case it calls the new functions

[hughperkins]

Hmmm, hey thats a pretty nice demo script you have there.

[szagoruyko]

there was a few things I had to patch in the script and in clnn, I can create issues on what needs to be done to support networks loaded with loadcaffe if you want

[hughperkins]

Yes please, sounds good :-)