diff options
author | Calvin <calvin@EESI> | 2013-02-19 08:15:55 -0500 |
---|---|---|
committer | Calvin <calvin@EESI> | 2013-02-19 08:15:55 -0500 |
commit | aac1c1beb22827feb8eedb518533aee84699750d (patch) | |
tree | dad681359ec80b014b0a30e97d0ca0fb8a5dcd8c /quikr.py | |
parent | c69b7e13895e8692afedb742bfdcc110bd982974 (diff) |
saving changes to work remotely
Diffstat (limited to 'quikr.py')
-rwxr-xr-x | quikr.py | 21 |
1 files changed, 14 insertions, 7 deletions
@@ -52,6 +52,7 @@ def main(): input_lambda = 10000 kmer = 6 xstar = quikr(args.fasta, trained_matrix_location, kmer, input_lambda) + np.savetxt("python.csv", xstar, delimiter=",") print xstar return 0 @@ -80,21 +81,27 @@ def quikr(input_fasta_location, trained_matrix_location, kmer, default_lambda): print "Detected Mac OS X" count_input = Popen(["count-osx", "-r", str(kmer), "-1", "-u", input_fasta_location], stdout=PIPE) + + + # load the output of our count program and form a probability vector from the counts counts = np.loadtxt(count_input.stdout) counts = counts / np.sum(counts) - counts = default_lambda * counts - trained_matrix = np.load(trained_matrix_location) - - # perform the non-negative least squares - # import pdb; pdb.set_trace() + + # loado our trained matrix + trained_matrix = np.load(trained_matrix_location) trained_matrix = np.rot90(trained_matrix) - xstar = scipy.optimize.nnls(trained_matrix, counts) - + + #form the k-mer sensing matrix + trained_matrix = trained_matrix * default_lambda; + + xstar, rnorm = scipy.optimize.nnls(trained_matrix, counts) + xstar = xstar / sum(xstar) + return xstar |