Python Programming for Biology: Bioinformatics and Beyond

NumPy implementation

With  NumPy  a  lot  of  the  pain  is  removed  from  matrix  manipulations,  and  in  particular

there are fewer explicit loops. In particular, the updateSomap function in the pure Python

implementation becomes just one line in the NumPy version. See

Chapter  24

 for  further

discussion of this implementation, since the code is just a repeat of what is there.

def selfOrganisingMap(inputs, spread, size, steps=1000):

---

nRows, nCols = size

vecLen = inputs.shape

somap = numpy.random.rand(nRows, nCols, vecLen)

influence = numpy.dstack([spread]*vecLen) # One for each feature

infWidth = (len(influence)-1) // 2

makeMesh = numpy.ix_ # Ugly

for s in range(steps):

decay = numpy.exp(-s/float(steps))

for vector in inputs:

diff = somap-vector

diff2 = diff*diff

dist2 = diff2.sum(axis=2)

index = dist2.argmin()

row = index // nRows

col = index % nRows

rows = [x % nRows for x in range(row-infWidth, row+1+infWidth)]

cols = [y % nCols for y in range(col-infWidth, col+1+infWidth)]

mesh = makeMesh(rows,cols)

somap[mesh] -= diff[mesh] * influence * decay

return somap