def doGrid(self, method=None):
if method is None:
feal = self.E.all_feal()
else:
self.space = method
feal = self.E.all_feal(force=True, method=method)
self.kdg = KDGrid(K=10, split=5)
self.kdg.insert([i[10:] for i in feal])
class Reweight:
def __init__(self, expr, space='cw'):
self.E = expr
self.space = 'cw'
self.last_method = None
self.kdg = None
self.hc = None
self.fealcov = None
self.covar = None
def doGrid(self, method=None):
if method is None:
feal = self.E.all_feal()
else:
self.space = method
feal = self.E.all_feal(force=True, method=method)
self.kdg = KDGrid(K=10, split=5)
self.kdg.insert([i[10:] for i in feal])
def doCovar(self, winsize=.1, slide=.05):
# Calculate Covariance matrices over all temporal data IAW given window param
self.covar = []
self.fealcov = []
for i, tr in self.E.trlist.items():
if i % 100 == 0:
print(i)
cov = dr.calc_covar(tr.xyz, winsize, 1, slide=slide)
if self.space in ['ds', 'dsw']:
X, C = dr.distance_space(tr), self.E.cent_ds
else:
X, C = tr.xyz, self.E.cent_c
wgt = self.E.cw if self.space in ['cw', 'dsw'] else [1,1,1,1,1]
rms = calc_rmsd(X, C, weights=wgt)
W = int(winsize * 1000)
S = int(slide * 1000)
feal = [np.mean([FL.feal.atemporal(i) for i in rms[st:st+W]], axis=0) for st in range(0, len(tr.xyz), S)]
for n in range(min(len(cov), len(feal))):
self.covar.append(cov[n])
self.fealcov.append(feal[n])
# Extract diagonals as a vector and solve Kernel PCA
def doTree(self, pc=4, ktype='rbf', leafsize=100, maxdepth=6, split='middle'):
diag = np.array([np.diag(i) for i in self.covar])
kpca = PCAKernel(pc, ktype)
kpca.solve(diag)
gdata = kpca.project(diag)
# Create KDTree over reduced-dim subspace (4-PCs)
gtree = KDTree(leafsize, maxdepth, gdata, split) # Or Max Gap
self.hc = gtree.getleaves()
def doHeatmap(self, title, DS=True):
# Save grid hcubes which actually have sufficient data points
nodeA_size = [len(i) for i in self.kdg.grid]
nodeA = [i for i, size in enumerate(nodeA_size) if size > 10]
nodeB = list(self.hc.keys())
idxB = {k: i for i, k in enumerate(nodeB)}
# For each HCube (among clustered temporal data):
for k in nodeB:
self.hc[k]['reweight'] = {}
for cov in self.hc[k]['elm']:
probe = self.kdg.toindex(self.fealcov[cov][10:])
if probe not in self.hc[k]['reweight']:
self.hc[k]['reweight'][probe] = 0
self.hc[k]['reweight'][probe] += 1
if probe not in nodeA:
nodeA.append(probe)
idxA = {k: i for i, k in enumerate(nodeA)}
# Map projection to edges (and subsequent heatmap chart)
edge = []
projmap = np.zeros(shape=(len(nodeA), len(nodeB)))
for kB in nodeB:
for kA, proj_cnt in self.hc[kB]['reweight'].items():
edge.append((kA, kB, proj_cnt))
A = idxA[kA]
B = idxB[kB]
projmap[A][B] = proj_cnt
xlabel = 'SOURCE: Temporal Windows (Covariance -> KPCA -> KDTree)'
ylabel = 'DEST: ATemporal Data (K-D Grid w/feature Landscape, 0-1..3-4 vals)'
Asize = [nodeA_size[i] for i in nodeA]
proj_total = [int(i) for i in np.sum(projmap, axis=1)]
Bsize = [self.hc[k]['count'] for k in nodeB]
alabel = ['#%04d/ %5d/ %3d' % x for x in zip(nodeA, Asize, proj_total)]
blabel = ['%6s/ %d' % x for x in zip(nodeB, Bsize)]
pmap_bal_row_norm = np.nan_to_num(projmap / np.linalg.norm(projmap, axis=-1)[:, np.newaxis]).T
pmap_bal_col_norm = np.nan_to_num(projmap.T / np.linalg.norm(projmap.T, axis=-1)[:, np.newaxis])
P.heatmap(projmap, alabel, blabel, title, ylabel=ylabel, xlabel=xlabel)
P.heatmap(pmap_bal_col_norm.T, alabel, blabel, title+'_NormCol', ylabel=ylabel, xlabel=xlabel)
P.heatmap(pmap_bal_row_norm.T, alabel, blabel, title+'_NormRow', ylabel=ylabel, xlabel=xlabel)
def drawLandscape(self, idxList, title):
feal = np.mean([self.E.feal_list[i] for i in idxList], axis=0)
var = np.std([self.E.feal_list[i] for i in idxList], axis=0)
P.feadist(feal, title, err=var, pcount=len(idxList), norm=10, negval=True)
def drawIndex(self, idxNum):
self.drawLandscape(self.kdg.index[idxNum], 'FeatLand_%d'%idxNum)
def drawHC(self, hcKey):
self.drawLandscape(self.hc[hcKey]['elm'], 'FeatLand_%s'%hcKey)
