def fit_patches(self, data, dq, parms, clip, psfs, noise_models):
"""
Fit the data given the psf_model
"""
flags = np.zeros_like(data, dtype=np.int)
fit_vars = np.zeros_like(data)
fit_masks = np.ones_like(data, dtype=np.bool)
for i in range(data.shape[0]):
fp, fv, ind = self.fit_single_patch(data[i], dq[i], psfs[i],
noise_models[i])
if i == 0:
fit_parms = np.zeros(data.shape[0], fp.shape[1])
if clip:
for j in range(parms.clip_iter):
# define model and clip variance
scaled_psf, bkg = self.construct_model_parts(fp[None, :], psfs)
model = scaled_psf + bkg
clip_var = parms.floor + parms.gain * np.abs(model)
clip_var += parms.q * scaled_psf ** 2.
# sigma clip
chi = np.zeros_like(data)
chi[ind] = np.abs(data[ind] - model[ind]) / np.sqrt(clip_var[ind])
condition = chi - parms.clip_tol
condition = (condition > 0).reshape(parms.patch_shape)
# redefine mask, grow and add to dq mask.
ind = 1 - ind.reshape(parms.patch_shape)
idx = grow_mask(condition)
flags[i] = ind.copy()
flags[i][idx] = 3
flags[i][condition] = 2
ind = np.ravel((ind == 0) & (idx == 0))
# refit
fp, fv, ind = self.fit_single_patch(data[i], dq[i], psfs[i],
noise_models[i], ind=ind)
fit_vars[i] = fv
fit_masks[i] = ind
fit_parms[i] = fp
return fit_parms, fit_vars, fit_masks
def fit_single_patch(data, psf, dq, parms, var=None):
"""
Fit a single patch, return the scale for the psf plus any
background parameters. Takes in flattened arrays for
data and psf.
"""
gain = parms.gain
floor = parms.floor
clip_parms = parms.clip_parms
background = parms.background
if var == None:
var = np.ones_like(data)
if background is None:
A = np.atleast_2d(psf).T
elif background == 'constant':
A = np.vstack((psf, np.ones_like(psf))).T
elif background == 'linear':
N = np.sqrt(psf.size).astype(np.int)
x, y = np.meshgrid(range(N), range(N))
A = np.vstack((psf, np.ones_like(psf),
x.ravel(), y.ravel())).T
else:
assert False, 'Background model not supported: %s' % background
ind = dq == 0
# fit the data using least squares
rh = np.dot(A[ind, :].T, data[ind] / var[ind])
try:
lh = np.linalg.inv(np.dot(A[ind, :].T, A[ind, :] / var[ind, None]))
fit_parms = np.dot(lh, rh)
except:
fit_parms = np.zeros(A.shape[1])
bkg = make_background(data, fit_parms, background)
# sigma clip if desired
if (clip_parms is not None) & (np.any(fit_parms != 0)):
Niter = clip_parms[0]
assert Niter == 1, 'Multiple rounds of clipping not supported.'
tol = clip_parms[1]
for i in range(Niter):
# define model and noise
scaled_psf = psf * fit_parms[0]
model = scaled_psf + bkg
if parms.plot_data:
parms.old_bkg = bkg
parms.old_model = model
model = model[ind]
scaled_psf = scaled_psf[ind]
var = floor + gain * np.abs(model) + (parms.q * scaled_psf) ** 2.
# sigma clip
chi = np.zeros_like(data)
chi[ind] = np.abs(data[ind] - model) / np.sqrt(var)
condition = chi - tol
condition = (condition > 0).reshape(parms.patch_shape)
# redefine mask, grow and add to dq mask.
ind = 1 - ind.reshape(parms.patch_shape)
idx = grow_mask(condition)
if parms.plot_data:
parms.flags = ind.copy()
parms.flags[idx] = 3
parms.flags[condition] = 2
ind = np.ravel((ind == 0) & (idx == 0))
# refit
rh = np.dot(A[ind, :].T, data[ind])
try:
lh = np.linalg.inv(np.dot(A[ind, :].T, A[ind, :]))
fit_parms = np.dot(lh, rh)
except:
fit_parms = np.zeros(A.shape[1])
bkg = make_background(data, fit_parms, background)
fit_vars = np.diagonal(lh)
return fit_parms, fit_vars, bkg, ind
def fig1(small=1.e-6):
"""
Figure showing examples of data patches
"""
Nexamples = 16
rows = 4
cols = 4
rsize = 4
csize = 4
shrink = 1.
ws, hs = 0.0, 0.2
fs = 24
# load data and dq
f = pf.open('../data/region/f160w_25_457_557_457_557_pixels.fits')
d = f[0].data
f.close()
f = pf.open('../data/region/f160w_25_457_557_457_557_dq.fits')
dq = f[0].data
f.close()
# toss data where dq is nonzero in center
center_pixel = (d.shape[1] - 1 ) / 2
ind = dq[:, center_pixel] == 0
d = d[ind]
dq = dq[ind]
# sort by maxima
mxs = d[:, center_pixel]
ind = np.argsort(mxs)
mxs = mxs[ind]
d = d[ind]
dq = dq[ind]
# get 7 other random patches
ind = [0, d.shape[0] - 1]
count = 0
while True:
new = np.random.randint(d.shape[0])
if new in ind:
continue
else:
ind.append(new)
if len(ind) == Nexamples:
break
# sort and reshape to 2D
ind = np.array(ind)
data = d[ind]
dq = dq[ind]
mxs = data[:, center_pixel]
ind = np.argsort(mxs)
patch_side_size = np.sqrt(data.shape[1]) # assumes square patches
dq = dq[ind].reshape(Nexamples, patch_side_size, patch_side_size)
data = data[ind].reshape(Nexamples, patch_side_size, patch_side_size)
# the figure
pl.gray()
f = pl.figure(figsize=(rows * rsize, cols * csize))
pl.subplots_adjust(wspace=ws, hspace=hs)
axes = [None] * cols
for i in range(Nexamples):
# patch pixels with dq != 0
bad_pix = np.where(dq[i] != 0)
for j in range(bad_pix[0].size):
ind = np.zeros_like(dq[i])
ind[bad_pix[0][j], bad_pix[1][j]] = 1
idx = grow_mask(ind)
fix = np.median(data[i][idx])
data[i][bad_pix[0][j], bad_pix[1][j]] = fix
mn = data[i].min()
mx = data[i].max()
kwargs = {'interpolation':'nearest', 'origin':'lower',
'norm':LogNorm(vmin=mn, vmax=mx)}
ax = pl.subplot(rows, cols, i + 1)
pl.imshow(data[i], **kwargs)
ax.set_xticks([])
ax.set_yticks([])
pl.title('[%0.1f, %0.0f]' % (mn, mx), fontsize=fs)
f.savefig('../plots/paper/fig1.png', bbox_inches='tight')
pl.close(f)
