Re: more graphs
Bill Yerazunis <[email protected]>
| Newsgroups | gmane.mail.spam.crm114 |
|---|---|
| Message-ID | <[email protected]> |
From: "Eric S. Johansson" <[email protected]> graphs from latest tests http://imagebin.ca/view/wmYBvhx.html there are three graphs in this one image. The default case, balanced, and variable limits. In the default case, if a good message scores less than three, it's retrained as good, if a bad message scores greater than -3, it's trained as bad. the balance case is basically the same as above except an equal number of good and bad messages are trained. For some reason these two graphs come out exactly the same and I'm not sure why. I need to look closer to see if I'm making a mistake. The last is scoring using variable limits. That is the red and green training limits start out -350 and 350 and work their way in words at a moderately quick rate. My interface code to CRM 114 crm114.py http://pastebin.ca/1025922 my test code reading messages from my repository and driving CRM 114 test framework http://pastebin.ca/1025929 my question is why am I getting such horrible score distribution. I'm even more concerned with your .css files. I see that they are well under 10% full- that's very strange for well-trained systems ( slot usage is asymptotic and dependent on the language and corpus, but I've never seen email corpora with this small a footprint). one other note: I've been running CRM 114 in virtual machines in production and, it does not behave well. With only 10 to 15 simultaneous scoring queries, each with a different user *css file set, I'm seeing load ever just climb over 12 and sometimes as high as 70. I'm not surprised. CRM114 is very cache sensitive; once you bust cache, it's gonna slow way down. It's amazing how fast it goes to hell. It gets worse during training processes which I have serialized. If a lot of training queues up, it's time to score messages seems to get a lot worse (there is a much longer queue for processing). I do some simple locking (shared access for reading and exclusive access for writing) because if I don't, and I start running parallel writes, we go right past hell to something else. That's strange. It should speed up when you enforce serialization... unless you're doing something like spinwaiting. Please tell me that you're not spinwaiting? Please? - Bill the CSS files are from the last test (variable limits). esj@tpblue-dev:~/projects/crm114-test$ cssutil esj/crm114/spam.css Sparse spectra file esj/crm114/spam.css statistics: Total available buckets : 1048577 Total buckets in use : 63701 Total in-use zero-count buckets : 0 Total buckets with value >= max : 0 Total hashed datums in file : 73716 Documents learned : 44 Features learned : 73717 Average datums per bucket : 1.16 Maximum length of overflow chain : 6 Average length of overflow chain : 1.10 Average packing density : 0.06 bin value 0 found 984873 times bin value 1 found 60546 times bin value 2 found 1791 times bin value 3 found 620 times bin value 4 found 182 times bin value 5 found 98 times bin value 6 found 117 times bin value 7 found 56 times bin value 8 found 32 times bin value 9 found 35 times bin value 10 found 37 times bin value 11 found 33 times bin value 12 found 30 times bin value 13 found 16 times bin value 14 found 7 times bin value 15 found 7 times bin value 16 found 1 times bin value 17 found 16 times bin value 18 found 1 times bin value 19 found 1 times bin value 20 found 2 times bin value 21 found 1 times bin value 22 found 1 times bin value 23 found 1 times bin value 24 found 1 times bin value 27 found 1 times bin value 28 found 1 times bin value 29 found 1 times bin value 31 found 1 times bin value 40 found 4 times bin value 41 found 6 times bin value 43 found 55 times bin value 44 found 1 times Options: Z n - zero bins at or below a value S n - subtract a constant from all bins D n - divide all bins by a constant R - rescan P - pack Q - quit >>> q Bye! esj@tpblue-dev:~/projects/crm114-test$ esj@tpblue-dev:~/projects/crm114-test$ cssutil esj/crm114/nonspam.css Sparse spectra file esj/crm114/nonspam.css statistics: Total available buckets : 1048577 Total buckets in use : 118231 Total in-use zero-count buckets : 0 Total buckets with value >= max : 0 Total hashed datums in file : 141185 Documents learned : 86 Features learned : 141186 Average datums per bucket : 1.19 Maximum length of overflow chain : 8 Average length of overflow chain : 1.19 Average packing density : 0.11 bin value 0 found 930343 times bin value 1 found 111779 times bin value 2 found 3584 times bin value 3 found 880 times bin value 4 found 561 times bin value 5 found 517 times bin value 6 found 213 times bin value 7 found 171 times bin value 8 found 75 times bin value 9 found 76 times bin value 10 found 32 times bin value 11 found 23 times bin value 12 found 45 times bin value 13 found 20 times bin value 14 found 30 times bin value 15 found 17 times bin value 16 found 18 times bin value 17 found 9 times bin value 18 found 16 times bin value 19 found 8 times bin value 20 found 8 times bin value 21 found 20 times bin value 22 found 10 times bin value 23 found 7 times bin value 24 found 7 times bin value 25 found 5 times bin value 27 found 1 times bin value 28 found 2 times bin value 29 found 2 times bin value 30 found 1 times bin value 31 found 2 times bin value 32 found 8 times bin value 33 found 1 times bin value 34 found 3 times bin value 35 found 1 times bin value 36 found 1 times bin value 40 found 1 times bin value 43 found 1 times bin value 44 found 1 times bin value 45 found 7 times bin value 51 found 1 times bin value 59 found 1 times bin value 63 found 1 times bin value 79 found 1 times bin value 84 found 2 times bin value 85 found 62 times bin value 86 found 1 times Options: Z n - zero bins at or below a value S n - subtract a constant from all bins D n - divide all bins by a constant R - rescan P - pack Q - quit >>> q Bye! esj@tpblue-dev:~/projects/crm114-test$ ------------------------------------------------------------------------- This SF.net email is sponsored by: Microsoft Defy all challenges. Microsoft(R) Visual Studio 2008. http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/ _______________________________________________ Crm114-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/crm114-general ------------------------------------------------------------------------- This SF.net email is sponsored by: Microsoft Defy all challenges. Microsoft(R) Visual Studio 2008. http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/