Result depends on order of input rttms about dover-lap HOT 17 CLOSED

Can you show me what command you ran?

from dover-lap.

Also, you should read this section: https://github.com/desh2608/dover-lap#combining-2-systems-with-dover-lap

from dover-lap.

Thank you for the answer.
I have not indeed read carefully the section about combining 2 systems. But I still don't understand why the order of input rttms may matter.
The commands I ran were:
dover-lap rttm_out1 rttm_in1 rttm_in2
and
dover-lap rttm_out2 rttm_in2 rttm_in1
And after running them I observed that rttm_out1 differs from rttm_out2 significantly.

In my experiments DER of rttm_in2 is lower than that of rttm_in1 by about 2%.
DER of rttm_out1 is similar to that of rttm_in1 (only a bit better but significantly worse than that of rttm_in2) while
DER of rttm_out2 is even better then that of rttm_in2 by about 1%.

from dover-lap.

Yes that is expected. We fall back to Hungarian mapping for the case of 2 RTTMs, which depends on the order of the provided inputs. You can think of it as the following situation: suppose you have 2 RTTMs (X and Y). DER computation between X and Y is not commutative, i.e., DER(X,Y) != DER(Y,X).

from dover-lap.

OK, I got it.
Answer please if the result is independent of the order for number of RTTMS >2 ? For each --label-mapping or just for greedy ?

from dover-lap.

In the latest dover-lap, we sort all RTTMs by default by average DER to other RTTMs, so for N>2, it should be independent of the order. Let me know if you find otherwise.

from dover-lap.

Thank you very much for clear explainations!

from dover-lap.

I have just checked doverlap from 3 input rttms and found the result doesn't depend of order of inputs for greedy mapping. But still depends for hungarian mapping. 6 different permutations lead to 6 different output RTTMs. Maybe sorting is applied just for greedy and order is preserved for hungarian?

from dover-lap.

Which version of dover-lap are you using?

from dover-lap.

I installed it several months ago. How can I get a version number?

from dover-lap.

I've found it. The version is 1.2.0. Will check latest soon

from dover-lap.

I have installed version 1.3.0 (BTW, I had to build it from sources and install since pip cannot find version upper than 1.2.0)
But I still obtain different results with different permutation of 3 input RTTMs and hungarian mapping
Although the difference seems to become less significant (the range of DER/JER fluctuations for 6 permutations is about 0.2% now, it was larger before)

from dover-lap.

If you can provide a minimum reproducible example, I will take a look later and try to fix.

from dover-lap.

In the meantime, I have created a release for v1.3: https://pypi.org/project/dover-lap/1.3.0/

from dover-lap.

doverlap.zip
Uploaded a small reproducible example. There are 3 files in zip, which I passed to dover-lap in different order (with --label-mapping=hungarian).
Hope you will look at them soon.

from dover-lap.

@kfmn can you please check if the above PR solves your problem?

from dover-lap.

The new version, which fixes this ordering issue, is available on PyPi: https://pypi.org/project/dover-lap/1.3.1/

from dover-lap.