def computeLikelihoodRealCDold(args):
"""
Args: (it's more convenient for multiprocessing)
args: a list of [R,s,h].
R: is a dataframe for which each row is a position and columns are allele frequencies.
ColumnsLevels= [REP, TIME] , IndexLevels=[CHROM,POS]
s: is selection strength
h: is overdominance
Returns:
a series containing likelihood of timeseries for the specific values of s and h.
"""
CD, E, s, h, regLambda = args
print CD.shape, s, h
if CD.shape[0] > 4 * 1e5:
numBatches = 5
idx = np.arange(CD.shape[0])
return pd.concat(
map(lambda x: computeLikelihoodRealCDold((CD.iloc[x], E, s, h, regLambda)),
np.array_split(idx, numBatches)))
powers = pd.Series(pd.Series(CD[r].columns).diff().values[1:] for r in range(3))
T = pd.read_pickle(utl.outpath + 'transition/real/S{:02.0f}.H{:02.0f}.df'.format(s * 100, h * 100))
likes = pd.Series(0, index=CD.index, name=(s, h))
for rep, df in CD.T.groupby(level=0):
alpha = E.loc[df.loc[(rep, 0)]]
for step, power in zip(range(1, df.shape[0]), powers[rep]):
alpha = alpha.values.dot(T.loc[power].values) * E.loc[df.loc[rep].iloc[step]]
likes += utl.vectorizedLog(alpha.mean(1).values)
return likes - regLambda * abs(s)
def computeLikelihoodRealBatch(args):
CD, E, T, powers = args
likes = pd.Series(0, index=CD.index)
for rep, df in CD.T.groupby(level=0):
alpha = E.iloc[df.loc[(rep, 0)]].values
for step, power in zip(range(1, df.shape[0]), powers[rep]):
alpha = alpha.dot(T.loc[power].values) * E.values[df.loc[rep].iloc[step].values]
#likes += utl.vectorizedLog(alpha.mean(1))
likes += utl.vectorizedLog(alpha.mean(1)) #it should be here
return likes
