dscompile

[olgert debas]

Hi, I just finished compiling eblearn on my machine (MAC OS X 10.6.8) and I am trying to prepare the data from mnist.

Running 

../../bin/dscompile /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_original/test -outdir /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_processed -dname test -dims 32x32x1 -kernelsz 4x4

I get I get an error that I cannot determine (sorry I am brand new to eblearn). Can you help?

many thanks,

olgert

odenas@bxlab05:$ ../../bin/dscompile /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_original/test -outdir /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_processed -dname test -dims 32x32x1 -kernelsz 4x4

_______________________________________________________________

             Dataset compiler for libeblearn library 

_______________________________________________________________

input parameters:

  dataset name: test

  dataset type: regular

  dataset precision: float

  images root directory: /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_original/test

  annotations directory: 

  ignored annotations directory: 

  output directory: /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_processed

  outputs: /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_processed/test_*.mat

  images pattern: .*[.](png|jpg|jpeg|PNG|JPG|JPEG|bmp|BMP|ppm|PPM|pnm|PNM|pgm|PGM|gif|GIF|mat|MAT)

  channels mode: RGB

  preprocessing: yes

  display: no

  display sleep: 0 ms.

  shuffling: no

  usepose: no

  useparts: no

  partsonly: no

  stereo: no

  max per class limitation: none

  max data limitation: none

  mexican_hat_size: 0

  preprocessing kernel size: [ 4x4 ]

  deformations: -1

  resizing method: bilinear

  output dimensions: [ 32x32x1 ]

  minimum input dimensions: <empty>

  maximum input dimensions: <empty>

  no padded: no

  scales: none

  laplacian pyramid: 0

_______________________________________________________________

Setting dataset name to: test

Setting preprocessing to: 4x4_bilinear32x32

Preprocessing produces 1 layers per sample.

Preprocessing modules:

0: resizepp module 4x4_bilinear32x32, resizing with method 0 to 32x32 while preserving aspect ratio, pp: rgb_to_rgb

Setting output directory to /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_processed

Setting temporary output directory to /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_processed

Setting target dimensions to 32x32x1

Setting saving mode to: dynset

Enabling individual sample saving.

Disabling saving sample layers separately.

Setting image pattern to .*[.](png|jpg|jpeg|PNG|JPG|JPEG|bmp|BMP|ppm|PPM|pnm|PNM|pgm|PGM|gif|GIF|mat|MAT)

Counting number of samples in /Users/odenas/projects/xpr_model/eblearn/demos/mnist/mnist_original/test/ ...

terminate called after throwing an instance of 'std::bad_cast'

  what():  std::bad_cast

Abort trap

[Pierre Sermanet]

Hi Olgert,

mnist is already a compiled dataset, you don't need to compile it.

My guess is that dscompile tries to load it as an image since it's looking for .mat images as well (see image pattern) but fails.

If you're just looking to preprocess mnist, just add a preprocessing module in front of your architecture in your conf.

For example add a rgb_to_rgb module, this will normalize the input globally (remove the mean of entire input and divide by the standard deviation of the entire input).

If you run into other problems, try the latest release (1.1) which is more stable than the current trunk: svn co https://eblearn.svn.sourceforge.net/svnroot/eblearn/tags/eblearn-1.1-r2350/ eblearn 

Pierre

--
 
 

[the_minion]

Hey Olgert,

Dscompile seems to be suffering from this bug in OSX in both the current SVN and in release 1.1

I shall fix it and get back to you.

Thank you for reporting.

[the_minion]

Olgert,

The bug seems to be when you compile eblearn with gcc on OSX. I don't see the same bug with clang on OSX and the dataset compiles succesfully. Can you try working with clang for now, till I look at why it fails with a gcc compile.

You can use the eblearn trunk that you have already, there is no need to revert to the 1.1 version.

To use clang as your compiler, instead of gcc, you have to

get "Command-line tools for XCode" from http://developer.apple.com/downloads
Install it.

in a terminal, go to eblearn/tools directory

export CC=/usr/bin/clang
export CXX=/usr/bin/clang++
make clean
make

Let me know if you have any issues.

--

Soumith

[olgert debas]

Hi Pierre,

thanks for the information. I had seen the compiled dataset, but was trying to walk through the tutorial. 

--olgert

[olgert debas]

I will, thanks a lot for looking into this! 

--olgert