Implement Octave binary `mat` format

[dktcoding]

Here's the implementation of GNU Octave binary mat format, the good:

• It works
• It's tested
• Multiple matrices per file works great (ideal for NN)

Some considerations (not actually good or bad):

• Only tested in Linux x64 and i386 (since that's what I have available)
• Tested with octave 4.0.2 x86_64-pc-linux-gnu
• GNU Octave path must be properly set before running the tests (I guess this might be a problem in Windows)
• Tests create files in the system's tmp dir, so it must exist and be rwx
• There are two magic magic numbers that I can't honestly explain, but I compare the hex dumps of multiple mat files created both in Linux 32-bits and Linux 64-bits and where always there (see comment in code)
• It uses 64-bit floats (as java)
• Matrices are always saved as dense matrices
• Numbers are saved in little-endian (not as Java, but as Octave's default)

Also, I took the liberty of change the target-source of the Netbeans project to 1.7 (since it's the minimum supported version, and It has a couple of small code advantages over 1.6).

[kronenthaler]

I was giving this a thought, i would like to implement something like ImageIO class. A class (factory) that allows to serialize a particular kind of object, keeping the object entirely agnostic of the different formats. For instance, a class MatrixIO, can have 2 methods, write(Matrix, format), and read(format) -> Matrix. Based on the format parameter (enum type?) a specific serializer is selected for the task. The advantages are that, the matrix class remains agnostic, no knowledge about multiple formats and more formats can be added without changing the Matrix itself.

Maybe just left in the Matrix class the toString method for debugging purposes, OpenOffice format can be also be extracted out.

Additional note: i want to integrate the project with travis-ci whenever i have time, it will allows to test it automatically, i will need to know what packages you need to install to test with octane :)

[dktcoding]

Sounds good, what about something like:

public class MatrixIO {
    public static enum Target {
        SERIAL, //default to jvm serialization
        CSV,
        TSV,
        OCTAVE_32,
        OCTAVE_64
    }

    public boolean write(String file, Matrix a) {}
    public boolean write(String file, Matrix a, Target t) {}
    public boolean write(String file, Map<String, Matrix> m, Target t) {
         //This one is needed (at least) for Octave since matrices need their own name
         //The other methods can have a default name, let's say 'a'
    }
}

As for packages, only octave is needed since I used built-in functions.

[kronenthaler]

I would just change the String file, to a OutputStream output, so we can output it on the console as well if needed. And the first method i will remove it too. I think you always should specify the type we want.

Later on we can include the read counterparts :)

On Thu, Dec 1, 2016 at 11:48 AM, Federico Vera notifications@github.com wrote:

Sounds good, what about something like:

public class MatrixIO { public static enum Target { SERIAL, //default to jvm serialization CSV, TSV, OCTAVE_32, OCTAVE_64 }

public boolean write(String file, Matrix a) {}
public boolean write(String file, Matrix a, Target t) {}
public boolean write(String file, Map<String, Matrix> m, Target t) {
     //This one is needed (at least) for Octave since matrices need their own name
     //The other methods can have a default name, let's say 'a'
}

}

As for packages, only octave is needed since I used built-in functions.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/kronenthaler/libai/pull/22#issuecomment-264140217, or mute the thread https://github.com/notifications/unsubscribe-auth/AAfVI3GMs1cIuKVcaD0CPuUjXRSRYzpxks5rDqYLgaJpZM4LBLCs .

--

.: Kronenthaler :.

kronenthaler@gmail.com | yahoo.com

[dktcoding]

Great! I'll send the commits later today

[dktcoding]

I also changed the NullPointerException to IllegalArgumentException since that's what ImageIO does now.