Added Cythonized interface for TSNE.

[sadaszewski]

No description provided.

[lvdmaaten]

Thanks, the current approach of writing data to a file and running an executable that reads those files is, indeed, super-clunky.

I'm wondering if/why we need the Cython dependency though? Wouldn't this be much easier solved by calling tsne_run via FFI? Something along these lines ought to work:

import ctypes
tsne_lib = ctypes.cdll.LoadLibrary('libtsne.so')
c_data = data.ctypes.data_as(ctypes.POINTER(ctypes.c_double))	
c_mappedX = mappedX.ctypes.data_as(ctypes.POINTER(ctypes.c_double))	
tsne_lib.tsne_run(	
        c_data,	
        ctypes.c_int(N),	
        ctypes.c_int(D),	
        c_mappedX,	
        ctypes.c_int(numDims),	
        ctypes.c_double(perplexity),	
        ctypes.c_double(theta),	
        ctypes.c_int(randomSeed),
        ctypes.c_int(skipRandomInit),	
        ctypes.c_int(maxIter),	
        ctypes.c_int(stopLyingIter),	
        ctypes.c_int(momSwitchIter),
)

This wouldn't require any additional dependencies. What do you think?

[sadaszewski]

Hi, thank you for a fast response. I think it is one of several alternatives. For reference: https://stackoverflow.com/a/5686873 . I found it quite straightforward with Cython. Also syntax sugar for specifying what type of NumPy arrays it takes in is quite nice. It makes perfect sense to provide ctypes interface with bhtsne and split Cython interface into a separate package or just stick to ctypes as long as correct handling of NumPy arrays is taken care of. Best, -- S.

…

On 19 March 2018 at 19:44, Laurens van der Maaten ***@***.***> wrote: Thanks, the current approach of writing data to a file and running an executable that reads those files is, indeed, super-clunky. I'm wondering if/why we need the Cython dependency though? Wouldn't this be much easier solved by calling tsne_run via FFI? Something along these lines ought to work: import ctypes tsne_lib = ctypes.cdll.LoadLibrary('libtsne.so') c_data = data.ctypes.data_as(ctypes.POINTER(ctypes.c_double)) c_mappedX = mappedX.ctypes.data_as(ctypes.POINTER(ctypes.c_double)) tsne_lib.tsne_run( c_data, ctypes.c_int(N), ctypes.c_int(D), c_mappedX, ctypes.c_int(numDims), ctypes.c_double(perplexity), ctypes.c_double(theta), ctypes.c_int(randomSeed), ctypes.c_int(skipRandomInit), ctypes.c_int(maxIter), ctypes.c_int(stopLyingIter), ctypes.c_int(momSwitchIter), ) This wouldn't require any additional dependencies. What do you think? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#69 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABUI3XaAknkg2juNl__vsAbP0eZA5yrtks5tf_ylgaJpZM4SuRYF> .