search

idiap / IBDiarization

C++ Implementation of the Information Bottleneck System
GNU General Public License v3.0
23 stars 11 forks source link

Toolkit Crash #13

Closed aishwaryjoshi31 closed 7 years ago

aishwaryjoshi31 commented 7 years ago

Toolkit crashed when I ran the experiment by including tdoa feature for diarization. Here is the snapshot of actual error. screenshot from 2017-04-24 17-14-04 Content of ".out file" AIB diarization started at 414731

MFCC: data/mfcc/IS1005a.feat weight MFCC 0.8 TDOA: data/tdoa/IS1005a.weat weight TDOA 0.2

Maximum Segment Duration 250

AIB is running with the following parameters

Maximum number of clusters possible: 10 Normalized Mutual Information threshold: 0.5 Beta value: 10 Number of threads: 1

Reading and Processing the scp file number of segments inside = 1716 Number of vectors = 39897 Feat file data/mfcc/IS1005a.feat of type 0 Reading feature file data/mfcc/IS1005a.feat frame dim: 12 num_vec = 39897 Attempting memory allocation. Memory successully allocated. Reading file ... all segments read Number of segments in the file = 1722 num_vec = 39897 idx = 0 Feat file data/tdoa/IS1005a.weat of type 2 Reading feature file data/tdoa/IS1005a.weat frame dim: 3148 num_vec = 39897 Attempting memory allocation.

mrsrikanth commented 7 years ago

Hello,

There seems to be an issue with the feature file data/tdoa/IS1005a.weat. Please make sure it is in the right format. I suspect type 2 is not a valid format for the weat features.

Thanks, Srikanth

aishwaryjoshi31 commented 7 years ago

Yes I changed the file from ".weat" to ".del" but still getting same error. I think the problem is because of memory allocation. If you can see the code for extract_feature.cpp, there it is reading HTK header for ".del" file and dimension of each frame comes out to be very large and allocating that much amount of memory for each frame is not feasible. So I suspect that something is wrong with frame_dimension or the format of ".del" file. I am attaching log file and the code snippet to get the clear picture of scenario. screenshot from 2017-04-27 16-32-16

mrsrikanth commented 7 years ago

Hello,

Is your feature dimensionality 3083? There are two possibilities here:

  1. Feature file is not in the right format. Please see the sample TDOA features in the toolkit.
  2. If this is indeed the feature you want to use, you do not have enough memory in your machine to run diarization with these features.

Srikanth

aishwaryjoshi31 commented 7 years ago

No size of feature dimension is only 8 but file interpretation is wrong. I exactly did what you said I first tried to check sample tdoa feature file but its content can't be read as garbage values are coming. I think the tdoa feature file in data folder is generated by some other toolkit rather than by BeamformIt).

mrsrikanth commented 7 years ago

The same toolkit was used. Are you sure that the .del file is in HTK format?

aishwaryjoshi31 commented 7 years ago

I don't think that BeamformIt tool generate in HTK format. Here is the content of file. There is no header nothing. image

mrsrikanth commented 7 years ago

Yes, that is expected. You need to convert this file to htk format. You need to choose the appropriate columns from this file to form the TDOA features. I suggest you to refer to their documentation.