Open GoogleCodeExporter opened 9 years ago
Could you post the output of the program up to that point? This behaviour maybe
due to a real bug or numerical precision problems.
The program is currently set to throw an error if the lower bound decreases. I
could alter that so that a warning is printed instead, but I need to confirm it
is just a precision problem.
Andrew
Original comment by AndrewJL...@gmail.com
on 11 Jan 2012 at 1:18
I tested 9 samples using joint-snv-mix-two train model for making parameter
files.
8 samples worked well.. One of them does not work.
I attach the output file of train error message.
seongjin
Original comment by psj...@gmail.com
on 11 Jan 2012 at 5:54
Attachments:
The output looks normal. Do you know what the average and max depth of coverage
are for that sample? One problem may be sites with extremely high depth.
If you need to process the files urgently before I get this fixed, you can use
the parameters from the end of that file to run the classify command.
Alternatively, if any of the cases you have analysed are a similar type of
cancer, you could also use the parameter file from them. The parameters from
other cases will probably be similar, as long as the same platform was used for
sequencing.
Original comment by AndrewJL...@gmail.com
on 12 Jan 2012 at 2:40
I get the same error. Is there a fix available yet?
Using SNVmix 0.7.5 on Mac OS 10.7 (Lion)
Only have limited output before the error (attached)
I can try to get coverage stats if wanted.
Original comment by dgoode.s...@gmail.com
on 30 Mar 2012 at 5:30
Attachments:
Sorry about the inactivity on this issue. I have recently had a simple idea on
how to fix this though.
It may be possible to avoid the issue by setting the --convergence_threshold
flag when training to a value above the default of 0.000001 to something like
0.01. This will hopefully cause the EM algorithm to stop before the numerical
issues kick in and the lower bound decreases.
In principle the estimates of the parameters will be a little worse, but I
think the difference in classification performance would be negligible.
Original comment by AndrewJL...@gmail.com
on 13 Jun 2012 at 8:08
The same issue.
I providing two logs: error and output.
Original comment by mik...@gmail.com
on 3 Jul 2012 at 1:03
Attachments:
Original issue reported on code.google.com by
psj...@gmail.com
on 10 Jan 2012 at 4:05