[mlpack] dataset for cnn

[Marcus Edel]

Hello Shangtong,

You could use a subset (or a subset of a subset) of the mnist dataset; mlpack
already runs some tests against the mnist_first250_training_4s_and_9s dataset.
The subset is selected to compare classes that aren't obviously separated which
doesn't mean that they couldn't be separated linearly in the original space.

Another option could be to implement e.g. the Capture Game task [1] and compare
the results against other implementations. The nice thing about the Capture Game
task is that you can easily scale the number of samples.

Hopefully this is helpful.

Thanks,
Marcus

[1] http://aura.abdn.ac.uk/bitstream/2164/748/1/konidaris02evolving.pdf <http://aura.abdn.ac.uk/bitstream/2164/748/1/konidaris02evolving.pdf>


> On 15 Feb 2015, at 03:05, Shangtong Zhang <zhangshangtong.cpp at qq.com> wrote:
> 
> Hi,
> 
> After implementing convolutional connection and pooling connection, 
> I want to do some tests to proof that my cnn does work.
> But I meet some problem.
> I build a small cnn and a toy input data and compare the output of the gradient and delta for each layer 
> with the value I calculate manually.
> The cnn’s behavior satisfies my expectation, in terms of numerical value.
> Then I want to test on real data. I choose MNIST dataset and LeNet1.
> But MNIST is too large, it has 60000 train data and 10000 test data. 
> And the network LeNet1 is also too complicate.
> My cnn doesn’t work in this test. 
> When training, the accuracy always stay unchanged and it’s very low.
> After thousands of iterations, my cnn tend to output the same value no matter how input varies.
> So I want to debug my cnn in a small dataset.
> Can anyone recommend me some small dataset for cnn? 
> It’s better if the data is 2-dim.
> 
> Thanks.
> 
> Shangtong Zhang,
> Third Year Undergraduate,
> School of Computer Science,
> Fudan University, China.
> _______________________________________________
> mlpack mailing list
> mlpack at cc.gatech.edu
> https://mailman.cc.gatech.edu/mailman/listinfo/mlpack

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20150215/f624eb84/attachment-0003.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 5136 bytes
Desc: not available
URL: <http://mailman.cc.gatech.edu/pipermail/mlpack/attachments/20150215/f624eb84/attachment-0003.bin>