hi, btgrahma.
I am so appreciated with your great work. But, i am a little confused with the following sentence in your paper:
"Using linear regression on just the CNN generated probability distributions, without any of the meta-data, also seems to work well."

linear regression is done based features probability distributions?
for example, you have 3 nets and repeat random transformation 3 times, in addition left-right eye, so you have 332*5=90 dimension of features for linear regression？

Hello, when I try to compile the project in Visual Studio 2013, I met an error. I found it quite similar with #2 , and I replaced that line with workers.emplace_back(std::thread(&BatchProducer::batchProducerThread,this,nThread)); but the error still exits.

1>C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include\functional(1148): error : no instance of overloaded function "std::_Pmd_wrap<_Pmd_t, _Rx, _Farg0>::operator() [with _Pmd_t=void (BatchProducer::*)(int), _Rx=void (int), _Farg0=BatchProducer]" matches the argument list
1>              argument types are: (BatchProducer *, int)
1>              object type is: std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>
1>            detected during:
1>              instantiation of "std::_Do_call_ret<_Forced, _Ret, std::decay<_Fun>::type, std::tuple<std::decay<_Types>::type...>, std::tuple<_Ftypes &...>, std::_Arg_idx<_Bindexes...>>::type std::_Bind<_Forced, _Ret, _Fun, _Types...>::_Do_call(std::tuple<_Ftypes &...>, std::_Arg_idx<_Bindexes...>) [with _Forced=false, _Ret=void, _Fun=std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>, _Types=<BatchProducer *, int>, _Ftypes=<>, _Bindexes=<0ULL, 1ULL>]" 
1>  (1137): here
1>              instantiation of "std::_Do_call_ret<_Forced, _Ret, std::decay<_Fun>::type, std::tuple<std::decay<_Types>::type...>, std::tuple<_Ftypes &...>, std::_Make_arg_idx<_Types...>::type>::type std::_Bind<_Forced, _Ret, _Fun, _Types...>::operator()(_Ftypes &&...) [with _Forced=false, _Ret=void, _Fun=std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>, _Types=<BatchProducer *, int>, _Ftypes=<>]" 
1>  C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include\thr/xthread(195): here
1>              instantiation of "unsigned int std::_LaunchPad<_Target>::_Run(std::_LaunchPad<_Target> *) [with _Target=std::_Bind<false, void, std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>, BatchProducer *, int>]" 
1>  C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include\thr/xthread(187): here
1>              instantiation of "unsigned int std::_LaunchPad<_Target>::_Go() [with _Target=std::_Bind<false, void, std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>, BatchProducer *, int>]" 
1>  C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include\thr/xthread(183): here
1>              instantiation of "std::_LaunchPad<_Target>::_LaunchPad(_Other &&) [with _Target=std::_Bind<false, void, std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>, BatchProducer *, int>, _Other=std::_Bind<false, void, std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>, BatchProducer *, int>]" 
1>  C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include\thr/xthread(205): here
1>              instantiation of "void std::_Launch(_Thrd_t *, _Target &&) [with _Target=std::_Bind<false, void, std::_Pmd_wrap<void (BatchProducer::*)(int), void (int), BatchProducer>, BatchProducer *, int>]" 
1>  C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\include\thread(49): here
1>              instantiation of "std::thread::thread(_Fn &&, _Args &&...) [with _Fn=void (BatchProducer::*)(int), _Args=<BatchProducer *, int &>]" 
1>  D:/Projects/diabetic/diabetic/BatchProducer.cu(123): here

Regards.

Dear Prof. Ben Warwick,

Nice work with this project!

I'm curious about running your spatially-sparse convolutional networks in the public cloud. Most public cloud options however don't have GPUs. So I was thinking of porting this to fallback to CPUs when GPUs aren't available.

Would welcome your opinion on this, and any suggestions on how best to approach this.

Thanks,

Samuel

PS: Obviously any decent solution I come up will be offered back in a PR

I am getting segmentation fault in the second epoch:
*1 Tesla K40c 11519MB Compute capability: 3.5
Sparse CNN - dimension=2 nInputFeatures=3 nClasses=10
0:Convolution 2^2x3->12
1:Learn 12->32 dropout=0 PReLU
2:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
3:Convolution 2^2x32->128
4:Learn 128->64 dropout=0 PReLU
5:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
6:Convolution 2^2x64->256
7:Learn 256->96 dropout=0 PReLU
8:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
9:Convolution 2^2x96->384
10:Learn 384->128 dropout=0 PReLU
11:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
12:Convolution 2^2x128->512
13:Learn 512->160 dropout=0 PReLU
14:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
15:Convolution 2^2x160->640
16:Learn 640->192 dropout=0 PReLU
17:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
18:Convolution 2^2x192->768
19:Learn 768->224 dropout=0 PReLU
20:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
21:Convolution 2^2x224->896
22:Learn 896->256 dropout=0 PReLU
23:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
24:Convolution 2^2x256->1024
25:Learn 1024->288 dropout=0 PReLU
26:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
27:Convolution 2^2x288->1152
28:Learn 1152->320 dropout=0 PReLU
29:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
30:Convolution 2^2x320->1280
31:Learn 1280->352 dropout=0 PReLU
32:Pseudorandom overlapping Fractional Max Pooling 1.25989 2
33:Convolution 2^2x352->1408
34:Learn 1408->384 dropout=0 PReLU
35:Pseudorandom overlapping Fractional Max Pooling 1.5 2
36:Convolution 2^2x384->1536
37:Learn 1536->416 dropout=0 PReLU
38:Learn 416->448 dropout=0 PReLU
39:Learn 448->10 dropout=0 Softmax Classification
(2,3) (4,5) (6,8) (9,11) (12,15) (16,20) (21,26) (27,34) (35,44) (45,57) (58,73) (74,93) Spatially sparse CNN: input size 94x94
epoch: 1 CIFAR-10 train set Mistakes:79.3582% NLL:2.11228 MegaMultiplyAdds/sample:820 time:93s GigaMultiplyAdds/s:437 rate:533/s
epoch: 2 Segmentation fault (core dumped)

Hi Benjamin,
The validation set file val_set is missing from the repository, it is not in Data/kaggleDiabeticRetinopathy/ of kaggle_Diabetic_Retinopathy_competition branch, only train_minus_val_set and test_set files are there. I know I can reconstruct it from trainLabels.csv and train_minus_val_set files but I would like to know if you can push it to the repository.
By the way, would you have the scripts to create a random train / validation partition (ie.: train_minus_val_set / val_set files) ?
Thank you

Trying to compile the source code on my Mac (OS X 10.10) with the latest CUDA 7.0 and Xcode drivers. I cannot get pass the compilation of BatchProducer.cu. The NVCC compiler complains when trying to full the thread vector (workers) in the BatchProducer constructor function. You have each thread binded to the creation of a batchProducerThread object. Unfortunately, I keep getting error that the thread is being binded to a deleted function. If change the thread target to a vector of threads defined explicitly outside the constructor, the everything works fine.

How you seen this before?

Hi,

I have a regression problem with sparse 1-dimensional input binary vectors and real valued output \in [0-1]. Should I be able to use SparseConvNet for it almost out of the box?

Thanks,
Ameen.

Installing seems to succeed, but than I get the following error upon import sparseconvnet as scn:

~/projects/SparseConv_sndbx/SparseConvNet/PyTorch/sparseconvnet/SCN/init.py in ()
1
2 from torch.utils.ffi import _wrap_function
----> 3 from ._SCN import lib as _lib, ffi as _ffi
4
5 all = []

ImportError: /home/bastiaan/projects/SparseConv_sndbx/SparseConvNet/PyTorch/sparseconvnet/SCN/_SCN.so: undefined symbol: _ZTISt12length_error

platform:
Ubuntu 16
Python 3.5
PyTorch 0.3.1

Any pointers anyone?

I am trying to apply SparseConvNet to ModelNet40 dataset (http://modelnet.cs.princeton.edu/ModelNet40.zip).

It seems that there is some memory leakage. After about 60 epochs the memory consumptino grows up from ~5Gb to ~15Gb,

My code can be found here https://github.com/yeahrmek/3DRetrieval (it is a fork of this project).
There modelNet.cpp is a copy of shrec2015.cpp modified to be used with ModelNet dataset.
I also modified OffSurfaceModelPicture class constructor to correctly load ModelNet OFF files (in some off files in ModelNet the first line is concatenated with the second)

Thank you for your work! I am studying 3d semantic segmentation. Since SparseConvNet is so useful.Is it possible to combine SparseConvNet with PointNet?