[mlpack] Out of memory error

[Ryan Curtin]

On Mon, Aug 20, 2018 at 10:52:15AM +0200, Anisa Llaveshi wrote:
> Greetings,
> 
> I have recently started using mlpack for a C++ application and I came
> across a problem that I haven't been able to solve. I am using Linear
> Regression to learn the parameters of a linear model. My training data is a
> vector of 1-dimensional points. It consists of a vector of type double
> (64-bit). I initialize the data points matrix from a std::vector structure
> (where I have the data) using this constructor: arma::mat(std::vector)
> Depending on the datasize of the dataset that I use to create the Linear
> Regression model I get this error:
> 
> 
> > error: arma::memory::acquire(): out of memory
> > terminate called after throwing an instance of 'std::bad_alloc'
> >   what():  std::bad_alloc
> >
> 
> I am running the application on a machine which has 250GB of memory. When I
> use 100k points (less then 1MB of data) I observe that ~28% of the memory
> is being used to build the model. When I increase this number to 160k
> points I observe that ~50% of the memory is being used and then the process
> is killed. When I increase it a bit more the above error is immediately
> thrown when trying to build the model.
> I was wondering whether it is normal for the model to consume this much
> memory for a small amount of data and if this is the case then what can one
> do to use a larger dataset?

Hi Anisa,

The memory used by a linear regression should be dxd, where d is the
number of dimensions in your data.  Remember that with mlpack, each
observation (or point) in your data should correspond to one
column---not one row---since Armadillo is column major. Here's some more
information:

https://www.mlpack.org/docs/mlpack-3.0.3/doxygen/matrices.html

So, my guess is that your matrix currently has N rows and 1 column,
causing mlpack to try and invert an NxN matrix (very large), but what
you actually want is a matrix with 1 row and N columns.

I hope this helps! Let me know if I can clarify anything.

Thanks,

Ryan

-- 
Ryan Curtin    | "Plot is a primitive vulgarity in literature."
ryan at ratml.org |   - Balph Eubank