HEAD | 4201050 | remove download stats for reasons I don't understand, it doesn't work. | 10 March 2021, 21:10:50 UTC |
refs/heads/buffer2 | c92c486 | Don't use stderr in ExternalProgramError because it’s empty now. | 19 April 2017, 09:13:03 UTC |
refs/heads/combined-speedup | e0c2f2d | remove pre-made task lists no evidence that they speed things up from empirical tests | 14 October 2016, 01:33:49 UTC |
refs/heads/develop | 47d2328 | fix bug in rclusterf this bug meant that if the median change was zero, we got stuck in an infinite loop. | 03 December 2015, 22:24:34 UTC |
refs/heads/feature-krmeans | 3378c1c | implement krmeans this is an idea for an algorithm in which the zero entropy sites are reassigned the entropy of their nearest physical non-zero entropy site in the alignment. Works fine so far. | 05 May 2016, 06:29:00 UTC |
refs/heads/feature/1kite-bugfix2 | ceaef5d | fixed bug in rcluster This is really fixed now. The issue was that the previous bug fix wasn’t bulletproof. It left the door open for a second bug, in which a single subset had an identical improvement score with >1 other subset. The new fix addresses this bug, as well as making sure that the original bug is fixed. | 25 August 2015, 00:14:20 UTC |
refs/heads/feature/DBSCAN | 7d0e1a2 | proto_DBSCAN | 19 August 2015, 13:10:55 UTC |
refs/heads/feature/complete_alignments | 5ba17e0 | improved user output for kmeans | 15 September 2015, 05:19:47 UTC |
refs/heads/feature/fabricated_subsets | 0eb9964 | fixed fabricated subset dealings at the end of the kmeans algorithm | 27 February 2015, 03:32:50 UTC |
refs/heads/feature/fastercluster | e394e8e | add two spaces | 13 November 2015, 02:46:55 UTC |
refs/heads/feature/fasttree | 1f628e8 | added write_fasta alignment function * FastTree requires interleaved phylip or fasta alignments. It is probably easier to write a fasta alignment so this function does that. | 02 September 2014, 15:29:48 UTC |
refs/heads/feature/fix-tests-pf2 | 55493fb | add init to make tests run | 26 February 2015, 06:45:24 UTC |
refs/heads/feature/garli_output | f9836c0 | Merge branch 'develop' into feature/garli_output | 30 April 2013, 01:14:54 UTC |
refs/heads/feature/greedy-speedy | 57715e4 | new version of greedy algorithm that borrows from the cluster algorithm, and is now a whole lot quicker and more efficient. | 07 September 2015, 22:25:35 UTC |
refs/heads/feature/importcheck | a0ea4df | some very minor changes | 18 September 2016, 23:35:19 UTC |
refs/heads/feature/iqtree | 77a3347 | first attempt at a whole bunch of IQtree model commandlines including R4-R8, R10, R12, R15, R20. will require some empirical tests to see which of the R’s are really needed. Ultimately, a progressive algorithm like that in IQtree (keep adding R cats until the AICc starts dropping) would be better. | 21 March 2017, 00:00:52 UTC |
refs/heads/feature/kmeans-manyparts | a4a5718 | make RAxML fall back on standard raxml with one CPU preparation for making the ML tree the default option | 25 July 2016, 22:54:19 UTC |
refs/heads/feature/krmeans2 | 6f760ad | new krmeans algorithm the previous version was naive. I reassigned invariant sites at every step, which just got the algorithm stuck early on. This version waits until the end of the kmeans algorithm to reassign sites, which is a much better idea. It appears to work well (in terms of AICc scores) on empirical datasets. | 12 May 2016, 07:58:51 UTC |
refs/heads/feature/lie-markov | d9ae785 | include category for lie markov models in models.cv These models have the attractive and possibly important property that you can multiply them together along branches and still have lie markov models. I don’t know of any evidence that inferences go wrong if you don’t use these models, but it’s possible. For a full description see e.g.: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4468350/ | 13 November 2015, 03:00:40 UTC |
refs/heads/feature/merge-little-subsets | 5851b59 | change scheme name when cleaning schemes so that it’s very obvious in the best_scheme.txt that you are using a cleaned scheme. | 11 November 2015, 22:02:54 UTC |
refs/heads/feature/model_csv | 7bc895f | implemented models.csv file, which is working in principle. | 26 February 2015, 06:38:55 UTC |
refs/heads/feature/morph_tiger | 0fd45b4 | Experimental morpho tiger rates This is a preliminary implementation to estimate tiger rates from a morphology alignment. | 25 November 2015, 19:22:30 UTC |
refs/heads/feature/morph_tiger_rates | 28a0cf1 | cleaned up print statements | 13 June 2016, 15:27:58 UTC |
refs/heads/feature/morphology | e0d9e80 | added dummy morphology models for phyml | 01 July 2013, 19:58:00 UTC |
refs/heads/feature/morphology2 | dcc88c3 | clean up model list checking still needs more work, but this is a good start | 21 November 2013, 21:31:59 UTC |
refs/heads/feature/new_clustering | a172fb3 | new relaxed clustering algorithm complete contains some much more efficient routines, including only making schemes once per step, and keeping a more efficient running tally of subset improvements. | 14 December 2013, 07:36:42 UTC |
refs/heads/feature/no-sleep | 50dd23a | remove sleep condition I suspect this is slowing us down a lot… | 29 September 2016, 04:12:00 UTC |
refs/heads/feature/phyml-external | 27137d2 | Update saved results files for latest phyml | 07 August 2015, 04:46:46 UTC |
refs/heads/feature/profiling | 2739914 | PEP8: Newline at end of file | 04 December 2013, 19:29:37 UTC |
refs/heads/feature/pytables | 61804e4 | Merge branch 'develop' into feature/pytables | 10 December 2013, 22:42:19 UTC |
refs/heads/feature/raxml-external | de2d12a | Fix raxml build | 07 August 2015, 07:41:11 UTC |
refs/heads/feature/test-tiger-arrays | d40f4d3 | Fix silly bugs after merge. | 09 March 2015, 06:47:49 UTC |
refs/heads/gh-pages | 0dfbde4 | github generated gh-pages branch | 04 July 2011, 10:14:21 UTC |
refs/heads/gui_test | c957068 | basic gui working | 30 June 2012, 03:57:23 UTC |
refs/heads/h5-bug | a874845 | ignore pyenv cruft | 03 October 2017, 19:55:50 UTC |
refs/heads/master | 4201050 | remove download stats for reasons I don't understand, it doesn't work. | 10 March 2021, 21:10:50 UTC |
refs/heads/paul_develop | d0cf8a8 | removed confusing log statement *log.info() statement read the number of sites from each codon position being split. This was for testing and, in reality, doesn’t work for most datasets. | 16 February 2015, 19:39:57 UTC |
refs/heads/release/1.1.0 | 27309ff | Handle expected failure of DNA_Clustering3 | 16 May 2013, 03:25:16 UTC |
refs/heads/speedup-threadpool | 3c5bfd0 | Create list of correct size for all tasks Should speed things up a little | 13 October 2016, 22:59:29 UTC |
refs/tags/h5-bugfix-1 | a874845 | ignore pyenv cruft | 03 October 2017, 19:55:50 UTC |
refs/tags/v0.9.1 | eccc508 | Use md5 to generate consistent length names | 07 March 2012, 11:43:38 UTC |
refs/tags/v2.0-pre1 | a554d7f | Better user output for kmeans | 14 August 2015, 11:13:37 UTC |
refs/tags/v2.0-pre2 | dba0794 | Merge pull request #63 from brettc/feature/1kitebugfix1 Feature/1kitebugfix1 | 16 August 2015, 02:09:03 UTC |
refs/tags/v2.0.0 | 41a5ef0 | update PF2 citation the ppr is now accepted | 22 November 2016, 04:50:43 UTC |
refs/tags/v2.0.0-pre10 | b6fcd69 | Merge pull request #80 from brettc/feature/fastercluster Feature/fastercluster we now have the search option rclusterf, which is a faster version of the rcluster algorithm. I do not yet know exactly how well it compares to rcluster, though it should be quite a bit faster in certain situations (especially where the number of models is << than the number of processors you have). | 13 November 2015, 05:14:57 UTC |
refs/tags/v2.0.0-pre11 | 47d2328 | fix bug in rclusterf this bug meant that if the median change was zero, we got stuck in an infinite loop. | 03 December 2015, 22:24:34 UTC |
refs/tags/v2.0.0-pre12 | 2d28e48 | remove unused test | 14 March 2016, 04:59:39 UTC |
refs/tags/v2.0.0-pre13 | a307ab8 | remove old debugging statement Embarrassing. Thanks to Ben Anderson for pointing this out. https://groups.google.com/forum/#!topic/partitionfinder/MSdcgxJ415w | 18 March 2016, 20:29:32 UTC |
refs/tags/v2.0.0-pre14 | acb84f8 | Merge pull request #104 from brettc/feature/morph Feature/morph | 31 May 2016, 22:39:13 UTC |
refs/tags/v2.0.0-pre15 | a06b857 | updated citation for PF2 | 18 September 2016, 23:41:24 UTC |
refs/tags/v2.0.0-pre16 | 7f70beb | fix windows bug reported here: https://groups.google.com/forum/#!topic/partitionfinder/4pAkDOHB5FM the bug was a hangover from the TIGER days. | 21 September 2016, 05:55:35 UTC |
refs/tags/v2.0.0-pre17 | e561bfa | update raxml version to https://github.com/stamatak/standard-RAxML/commit/5d9558ac18ddb2c69dd75a 9dc971bcf541bbfeb2 | 22 September 2016, 06:29:18 UTC |
refs/tags/v2.0.0-pre3 | e7529ea | updated gitignore | 04 May 2015, 07:21:07 UTC |
refs/tags/v2.0.0-pre4 | 97b68ef | updated manual contents | 25 August 2015, 03:55:24 UTC |
refs/tags/v2.0.0-pre5 | 8bd784c | changed user output for cluster | 25 August 2015, 05:17:09 UTC |
refs/tags/v2.0.0-pre6 | fb4fcd3 | remove -U option for RAxML it might be causing issues, and won’t work with morphology data. | 28 August 2015, 23:52:38 UTC |
refs/tags/v2.0.0-pre7 | b106624 | update kmeans test since we now disallow multiple subsets as input | 12 September 2015, 07:55:45 UTC |
refs/tags/v2.0.0-pre8 | 83be0bd | Merge pull request #70 from brettc/feature/complete_alignments Feature/complete alignments | 15 September 2015, 05:22:05 UTC |
refs/tags/v2.0.0-pre9 | a2d3b33 | updated manual added in —all-states and —min-subset-size | 02 October 2015, 04:17:43 UTC |
refs/tags/v2.1.0 | 19d7fe4 | Disable k-means for all but morphology #Why? A paper came out yesterday (http://www.sciencedirect.com/science/article/pii/S1055790316302780) that raises some serious concerns about the k-means algorithm, suggesting that it might lead to bad inferences on empirical datasets. I had spoken to the authors of the paper when they were revising it, but wasn’t aware until yesterday of the details of the problems they’d uncovered. Given how odd the inferences from k-means look, we decided to disable the method for all but morphological analyses (see below). # But there was a warning before, why disable it now? Our previous concerns came from our own realisation about one aspect of the method (that it lumps together all invariant sites) and some concerns raised by folks in Brian Moore’s lab this year. Specifically, we put in the warning when we learned that some simulated datasets that were analysed with k-means partitioning schemes led to bad inferences. I was hopeful that these simulations would be corner cases, and/or that one aspect of the simulations where k-means was misleading (that you got implausibly long trees) would mean that it would be trivial to diagnose cases in which there were issues. In addition, we had tried the method on lots of empirical datasets, and never seen any issues. Indeed, on at least one dataset the k-means tree seemed much more reasonable than trees we were getting from other methods. (I note that Brian Moore and co were less optimistic, and suggested from the start that we should consider disabling the method.) The empirical results in the recent paper suggest otherwise, and suggest that the best we can say of k-means for now is: ‘you should try other methods too, and if the methods disagree, we’d suggest ignoring the k-means tree'. On this basis, there seems little point keeping k-means as an available method: if you can't trust the reuslts, why bother. # I liked/used it, what should I do? Use standard methods, e.g. partitioning by codon position and locus, instead. Even better (if your dataset is small enough) use the automatic partitioning solutions in BEAST2 and/or MrBayes (google AutoParts). If you have used k-means to make an inference, it would be worthwhile to check that the inference is robust when you use a standard partitioning scheme too. # What’s the problem? We don’t know for sure, but it’s likely to be related to the fact that k-means separates out all invariant sites into a single subset. I presented on this at SMBE in July this year, but this has a couple of downstream effects. First, it makes AIC/AICc/BIC scores look really great, because when you have all the invariant sites together, you can estimate a rate of zero and get likelihoods of 1 for all of those sites. That’s a bit silly, and something I wish we’d realised earlier. Second, and more seriously for inference, putting all the invariant sites into one subset means that the other subsets have NO invariant sites. If you then analyse these without a model that accounts for this (e.g. with some kind of ascertainment bias) this is likely to mess with estimate of rates, branch lengths, and topologies. It’s not totally obvious yet how common the problem is, but now we’ve seen it in simulated and empirical datasets, it seems wise to can the method until we completely understand the problem and can fix it. # Are you going to fix it? We're working on it, but it will take a while. Apart from anything else, we are going to be exceptionally cautious in proposing more new methods related to this one. # But why is it still available for morphology? We’ve kept it in there for morphological datasets as an experimental method, and provide lots of warnings when you run the code and in the output that it’s experimental, untested etc. We did this because morphological datasets are different: they tend to have no invariant sites, and people tend to use models that correct for ascertainment bias. Because of that, it seems worthwhile to leave it in. We are working on testing it as exhaustively as possible for these datasets. # I want to use it anyway If you want to use it for empirical inferences, just don’t. But if you want to use it to try and figure out why it doesn’t work, and how you might improve it, then all you need to do is edit out the line that raises the error. # I have questions… Post on the google group or raise an issue on GitHub. | 02 December 2016, 05:08:47 UTC |
refs/tags/v2.1.1 | 63d5af1 | bump version number | 06 December 2016, 01:45:13 UTC |