turinglife / cuda-convnet Goto Github PK

When using startFeatureWriter to get out a feature map, I cannot seem to 
visualise the data properly. It seems there is some sort of channel 
interleaving, which is different depending on the layers. Is there anything I 
am missing to be able to visualise the output featuremap of a layer (even a 
simple layer such as resize or gaussian blur)?

It seems that the number of filters must be a multiple of 16, any solution to 
this limitation?

Hi Alex,

Could you confirm that this slightly strange-looking line from filter_act.cu is 
correct?

assert(paddingStart <= 0 && paddingStart + (numModules-1)*moduleStride + 
filterSize >= imgSize);

In particular, why are you multiplying numModules (which the square of 
numModulesX) by the stride and adding filterSize (which is not the square, but 
the side-length).

If you're sure, I trust you, but if you could additionally lend some intuition 
for why I'd appreciate it.

There is a comment in the code but I still don't get it:
// These routines don't handle the case when only part of the image is visited 
in the convolution

Thanks,
- James

It seems like Layer with weights class lacks a destructor. I wonder why it is 
so? Won't it cause memory leaks (for example, biases are never deleted)?

fprop(NVMatrixV&) does not remember forward activities, so backprop will fail. 
Currently backprop relies on forward activities being provided by layers below.

What steps will reproduce the problem?
1. use a big image like 512 x 512
2. put lots of filters (like 64)
3. have lots of color channels (again 64?)

What is the expected output? What do you see instead?
I expect a big filtered image, but instead it crashes. 

The blocks are defined such that blocks.y > (2^16) so CUDA refuses to launch 
the kernel.

I'm not sure I understand how to set the number of modules when doing a normal 
convolution, but it seems that an outer loop is required. The trouble with an 
outer loop is that the data is arranged in such a way that it is impossible to 
apply just a fraction of the  filters, or to process just some of each image. 
The data arrangement makes it natural to process just some of the image 
channels... but the color channels don't come into the blocking structure.

Basically... can I use this kernel to perform big convolutions?

Add support for local non-convolutional layers.

What steps will reproduce the problem?
1. 1000 images of the size 256x256 into a batch
2. modification of convdata.py
3. Try to train the network

What is the expected output? What do you see instead?
It's working with the size 32x32, 64x64 but if the images are bigger than that 
I have a "device memory allocation error". I think it comes from nvmatrix.cu, 
the cublasallocation is not working. 


What version of the product are you using? On what operating system?
I have a gpu GTX 590.  

Since the CUDA 5 I have had no success in compiling the cuda convnet code. 
Are there any plans to make it compatible with CUDA 5?

I can't download the code...
can anybody give me a new link? thanks

What steps will reproduce the problem?
1. train a model
2. multiview test the model and --test-out=1
3.

What is the expected output? What do you see instead?
probs matrix of multiview tested result.
All zero matrix

What version of the product are you using? On what operating system?
latest version

Please provide any additional information below.
Is --test-out function not yet developed? Since I saw the part of writing probs 
matrix is commented. Thanks

What steps will reproduce the problem?
1. vs2010 windows 7 64 bit
2. compile the code
3. ptxas fatal   : Memory allocation failure from logs
I have nvidia GTS 450 gpu: will GTS 450 gpu work ? or i need GTX gpu only?

i have attached pyconvnet.log file 

i have nvidia GTS 450 gpu, is this an error due to gpu gts , do i need to 
install gtx gpu?

The network training is fine without adding any contrast normalization layer 
(all types), but ones add the contrast normalization layers, after several 
iterations the net gets nan values.  I tried different values of the size, 
scale and pow values, and tried to place the layer before and after pooling 
layer. 

What steps will reproduce the problem?
1. compiler error,
2.
3.

What is the expected output? What do you see instead?
when i compile usr ./build.sh will broke see this error

obj/x86_64/release/src/util.cu.o: In function `pyDictGetMatrix(_object*, char 
const*)':
tmpxft_0000033f_00000000-3_util.cudafe1.cpp:(.text+0x27a): undefined reference 
to `Matrix::Matrix(PyArrayObject const*)'
obj/x86_64/release/src/util.cu.o: In function `getMatrixV(_object*)':
tmpxft_0000033f_00000000-3_util.cudafe1.cpp:(.text+0x3fe): undefined reference 
to `Matrix::Matrix(PyArrayObject const*)'

What version of the product are you using? On what operating system?
centos 6.3
gcc 4.4.6-4
cuda 5.0

Please provide any additional information below.

The result would be incorrect if the target is same as the first operand. The 
target==this version would require this to be of column major. I modified it so 
that this requirement is no longer needed:

void NVMatrix::rightMult(const NVMatrix &b, float scaleAB, NVMatrix &target) 
const {
    assert(isContiguous() && b.isContiguous() && target.isContiguous());
//    assert(&target != &b);
    assert(_numCols == b.getNumRows());
    if(&target != this) {
        target.resize(_numRows, b.getNumCols());
        //target.setTrans(true); // default column major
    }
    assert(target.getNumRows() == _numRows);
    assert(target.getNumCols() == b.getNumCols());
    if(_numRows % 64 != 0 || _numCols % 64 != 0 || b.getNumCols() % 64 != 0) {
        WARN("Matrix dimensions not divisible by 64 -- cublasSgemm performance may suffer.");
    }
    cublasSgemm(getTransChar(), b.getTransChar(), _numRows, b.getNumCols(), _numCols,
                scaleAB, _devData, getLeadingDim(), b.getDevData(), b.getLeadingDim(),
                0, target.getDevData(), getNumRows());
    target.setTrans(true); // added isTrans specification
    checkCublasError("cublasSgemm failed");
//    cudaThreadSynchronize();
}