def gen_initial_guess(observed_counts, one_sided_prf, Nguess, randscale=1e-3, reg='tv'):
"""
generate a list of initial guesses based on the RL deconvolution
"""
M = len(one_sided_prf)
symmetric_prf = np.r_[np.zeros(M-1), one_sided_prf]
Ndim = len(observed_counts) + M - 1
# pad first to avoid end effects
pad0 = np.ones(2*M)*observed_counts[0]
pad1 = np.ones(M)*observed_counts[-1]
observed_counts_padded = np.r_[pad0, observed_counts, pad1]
initial_guess = util.deconvlucy1d(observed_counts_padded, symmetric_prf,
iterations=1000, reg=reg)
## smooth the initial guess
#from micromet.util import smooth
#initial_guess = smooth(initial_guess, 11)
initial_guess = initial_guess[M+1:-M]
# starting locations for each walker are in a small range around the
# Richardson-Lucy starting point
p0 = [initial_guess*2**(np.random.randn(Ndim)*randscale) for
ii in range(Nguess)]
#p0 = [initial_guess + np.random.randn(Ndim)*randscale for
# ii in range(Nguess)]
return p0
int(itm[1]['month']),
int(itm[1]['dom'])),
itm[1]['time']) for itm in df_glb.iterrows()]
df_glb.index = time
#clean up negative values
df_glb.lld[df_glb.lld<0] = 500
return df_glb
df_glb = load_glb()
## exclude the calibration
#df_glb = df_glb.iloc[48*2:-12*2]
#rl deconvolution
df_glb['lld_rltv'] = util.deconvlucy1d(df_glb.lld, prf30_symmetric, reg='tv', iterations=1000)
df_glb['lld_rl'] = util.deconvlucy1d(df_glb.lld, prf30_symmetric, reg=None, iterations=1000)
f, ax = plt.subplots()
df_glb[['lld', 'lld_rl', 'lld_rltv']].plot(ax=ax)
# look at detailed diagnostics
if True:
#dftmp = df_glb.head(24*2)
dftmp = df_glb.iloc[48*2:48*3+1]
observed_data = dftmp.lld.values
# check out the initial guess
ig = util_emcee.gen_initial_guess(dftmp.lld.values, prf30, Nguess=1,
