def retrain(models, data, calibdir, finaldir, nfold=1, ntrial=1):
kangaroo.globals.flags.push("disable_softmax",True)
for trial in range(10):
Z = kangaroo.predict(data, finaldir, finaldir)["bound"]
I0 = [i for i in range(len(data)) if data.Y[i] < .5]
I1 = [i for i in range(len(data)) if data.Y[i] > .5]
#threshold = np.mean(Z[I1].ravel()) - 4*np.std(Z[I1].ravel())
#threshold = np.percentile(Z[I0].ravel(), 95)
threshold = np.percentile(Z[I1].ravel(), 1)
numshuffle = 0
for i in I0:
if Z[i] > threshold:
data.sequences[i][0] = doublet_shuffle(data.sequences[i][0])
data.X_seq[i] = data.sequences[i][0]
numshuffle += 1
numshuffle_pct = float(numshuffle)/len(I0)*100
print "retrain trial %d: had to shuffle %.1f%% unbound sequences" % (trial, numshuffle_pct)
if numshuffle <= .0001*len(I0):
break
kangaroo.globals.flags.pop("disable_softmax")
kangaroo.train(models, data, calibdir, finaldir, nfold=nfold, ntrial=ntrial)
return
def train_without_outliers(cfgs, data, calibdir, outdir, ntrial=1, auxfilter=None):
# Same as kangaroo.train, except this version trains twice: once with all
# the data, and again with worst outliers 'removed'.
# Step 1. Train with all training data
#kangaroo.train(cfgs, data, calibdir, outdir, nfold=1, ntrial=ntrial, auxfilter=auxfilter)
# Step 2. Load predictions on the training data, and remove%
# the examples that were misclassified the worst.
if not isinstance(data,dict):
data = { id : data.astargets([id]) for id in data.targetnames }
kangaroo.globals.flags.push("disable_softmax",True)
pred = kangaroo.predict(data, outdir, outdir)
for targetname in sorted(data.targetnames):
for name, Z in pred.iteritems():
Y = data[name].Y.ravel()
allidx = []
allseq = set()
roworder = np.argsort(-Z.ravel())
def save_metrics(data, groupname, outdir, modeldir=None):
if modeldir is None:
modeldir = outdir
if not isinstance(data,dict):
data = { id : data.astargets([id]) for id in data.targetnames }
pred = kangaroo.predict(data, modeldir, outdir)
for targetname in data.keys():
z = pred[targetname].ravel()
y = data[targetname].Y.ravel()
rowidx = data[targetname].rowidx
_update_metrics(outdir, targetname, groupname, rowidx, z, y)
check_repeat_correlation = False
if check_repeat_correlation:
for targetname in data.keys():
z = pred[targetname].ravel()
y = data[targetname].Y.ravel()
s = [data[targetname].X_seq[i] for i in range(len(data[targetname]))]
# Only get the bound examples.
z = [s[i] for i in range(len(data[targetname])) if y[i] != 0]
s = [s[i] for i in range(len(data[targetname])) if y[i] != 0]
class revcomp_str(object):
def __init__(self, s): self.s = s;
def __eq__(self, other): return self.s == other.s or self.s == revcomp(other.s)
# For each percentile, count the number of repeats
print targetname
order = np.argsort(z)
for denom in reversed([2,8,32,128,512,2048]):
top_s = [s[i] for i in order[len(order)//denom:]]
top_unique = set([revcomp(top_s[i]) for i in range(len(top_s))])
nall = float(len(top_s))
nunique = float(len(top_unique))
percent_repeat = 100*(nall - nunique) / nall
print " top %.4f%% (n=%d)\t=> %.2f%% repeats" % (100./denom, nall, percent_repeat)
def save_featuremaps(data, modeldir, outdir, maxrows=1000000):
if not isinstance(data,dict):
data = { id : data.astargets([id]) for id in data.targetnames }
disable_softmax()
pred = kangaroo.predict(data, modeldir, outdir)
# Find a list of rows, in order of decreasing confidence, and with all
# duplicate sequences deleted
for name, Z in pred.iteritems():
Y = data[name].Y.ravel()
allidx = []
allseq = set()
roworder = np.argsort(-Z.ravel())
for i in range(len(roworder)):
s = data[name].X_seq[roworder[i]]
#if s in allseq:
# continue
allidx.append(roworder[i])
allseq.add(s)
if maxrows and len(allidx) >= maxrows:
break
# Now actually dump the featuremaps for all the rows specified
print "Generating feature maps...",
datasub = data[name][allidx]
kangaroo.globals.set_multiprocessing(False) # Needed to get back collected values in globals; uugh
if "reverse_complement" in kangaroo.globals.flags:
kangaroo.globals.flags.push("collect_Zmask",True)
kangaroo.globals.flags.push("collect_featuremaps",True)
kangaroo.globals.flags.push("disable_relu",True)
kangaroo.predict(datasub, modeldir, outdir)
kangaroo.globals.flags.pop("disable_relu")
fmaps = kangaroo.globals.flags.pop("collect_featuremaps")
Zmask = kangaroo.globals.flags.pop("collect_Zmask").ravel() if "reverse_complement" in kangaroo.globals.flags else None
kangaroo.globals.set_multiprocessing(True)
seqs = []
for i in range(len(datasub)):
s = datasub.sequences[i][0]
seqs.append(s)
if "reverse_complement" in kangaroo.globals.flags:
seqs.append(revcomp(s))
if "reverse_complement" in kangaroo.globals.flags:
# Remove the strand that was not used for prediction
fmaps = [fmaps[i] for i in range(len(fmaps)) if Zmask[i]]
seqs = [seqs[i] for i in range(len(seqs)) if Zmask[i]]
filter_len = fmaps[0].shape[1] - len(seqs[0]) + 1
# Make tuples of (sequence, max_value, max_index) for each featuremap
pfmargs = []
for k in range(fmaps[0].shape[0]):
pfmarg = [(seqs[i], float(np.max(fmaps[i][k])), int(np.argmax(fmaps[i][k]))) for i in range(len(seqs))]
pfmargs.append(pfmarg)
pfmargs = np.array(pfmargs, dtype='a%d,f4,i4' % max(len(seqs[i]) for i in range(len(seqs))))
print "done"
#np.savez_compressed(outdir + "/%s.pfm_info.npz"%(name), pfmargs=pfmargs)
# Compute PFMs from the pfmargs array
print "Computing PFMs for %s..." % name,
pfms, ic, kl_dist, pval, counts = compute_pfms(pfmargs, filter_len=filter_len, num_permut=500, rev_comp=False)
print "done"
makepath(outdir)
with open(outdir + "/%s.pfms.pkl"%(name),"wb") as f:
cPickle.dump({"pfms" : pfms, "ic" : ic, "kl_dist" : kl_dist, "pval" : pval, "counts" : counts}, f)
enable_softmax()
