def geweke_source_params(srcs, imgs, Niter = 10):
e_samps = np.zeros((Niter, len(imgs))) # sky term
t_samps = np.zeros((Niter, len(srcs))) # source temp
b_samps = np.zeros((Niter, len(srcs))) # source brightness
u_samps = np.zeros((Niter, len(srcs), 2)) # source location
ll_samps = np.zeros(Niter)
prev_ll = celeste_likelihood_multi_image(srcs, imgs)
print prev_ll
for n in range(Niter):
if n%1==0:
print "===== iter %d of %d (curr_ll = %2.2f)==="%(n, Niter, prev_ll)
if n%100==0:
save_samples(e_samps, t_samps, b_samps, u_samps, ll_samps, fname='star_samples_teff_seed.bin')
# run a gibbs step
if n > 0:
celeste_gibbs_sample(srcs, imgs, subiter=1, verbose=False, debug=False)
# save current likelihood and samples
ll_samps[n] = celeste_likelihood_multi_image(srcs, imgs)
prev_ll = ll_samps[n]
for n_img in range(len(imgs)):
e_samps[n, n_img] = imgs[n_img].epsilon
for s in range(len(srcs)):
t_samps[n, s] = srcs[s].t
b_samps[n, s] = srcs[s].b
u_samps[n, s, :] = srcs[s].u
return t_samps, b_samps, u_samps, e_samps, ll_samps
def geweke_source_params(srcs, imgs, Niter = 10):
e_samps = np.zeros((Niter, len(imgs))) # sky term
t_samps = np.zeros((Niter, len(srcs))) # source temp
b_samps = np.zeros((Niter, len(srcs))) # source brightness
u_samps = np.zeros((Niter, len(srcs), 2)) # source location
ll_samps = np.zeros(Niter)
prev_ll = celeste_likelihood_multi_image(srcs, imgs)
print prev_ll
for n in range(Niter):
if n%1==0:
print "===== iter %d of %d (curr_ll = %2.2f)==="%(n, Niter, prev_ll)
if n%100==0:
save_samples(e_samps, t_samps, b_samps, u_samps, ll_samps, fname='star_samples_teff_seed.bin')
# run a gibbs step
if n > 0:
celeste_gibbs_sample(srcs, imgs, subiter=1, verbose=False, debug=False)
# save current likelihood and samples
ll_samps[n] = celeste_likelihood_multi_image(srcs, imgs)
prev_ll = ll_samps[n]
for n_img in range(len(imgs)):
e_samps[n, n_img] = imgs[n_img].epsilon
for s in range(len(srcs)):
t_samps[n, s] = srcs[s].t
b_samps[n, s] = srcs[s].b
u_samps[n, s, :] = srcs[s].u
return t_samps, b_samps, u_samps, e_samps, ll_samps
def testBirth():
np.random.seed(42)
##
## Generate some fake sources using real image data (PSF and stuff)
##
cat_glob = glob('../data/stamp_catalog/cat*.fits')[1:2]
srcs, imgs, teff_catalog, us = load_imgs_and_catalog(cat_glob)
print "initialized with %d sources and %d images"%(len(srcs), len(imgs))
## pixel location of source in image 0
u_pixel = imgs[0].equa2pixel(srcs[0].u)
## generate fake sources information
t_gt = np.array([5500]) # synthetic temperatures
b_gt = np.array([5e-10]) # synthetic brightnesses
gt_srcs = []
for s in range(1):
u_equa = imgs[0].pixel2equa(u_pixel)
src_s = celeste.PointSrcParams(u = u_equa, b = b_gt[s], t = t_gt[s])
gt_srcs.append(src_s)
print "Source at ", gt_srcs[0].u
print "Brightness ", gt_srcs[0].b
# re-generate images using these source params
for img in imgs:
mimg = celeste.gen_model_image(gt_srcs, img)
img.nelec = np.random.poisson(mimg)
# check out the first image
#compare_to_model(gt_srcs, imgs[0])
#plt.show()
np.random.seed(42)
rand = np.random.RandomState()
Niter = 1
e_samps = np.zeros((Niter, len(imgs))) # num images is fixed
ll_samps = np.zeros(Niter)
srcs = []
post_samps = [srcs] #
for iter_n in range(Niter):
print "Doing iteration", iter_n
# fresh new universe model
srcs = copy.deepcopy(srcs)
srcs = birthStar(srcs, imgs, rand=rand)
# keep track of universe model and image specific noise params
post_samps.append(srcs)
for n_img in range(len(imgs)):
e_samps[iter_n, n_img] = imgs[n_img].epsilon
ll_samps[iter_n] = celeste.celeste_likelihood_multi_image(srcs, imgs)
print "After iteration %s: %s, cat len %s" % (iter_n, ll_samps[iter_n], len(srcs))
print "Location: ", [src.u for src in srcs]
print "Brightness: ", [src.b for src in srcs]
def testDeath():
np.random.seed(42)
##
## Generate some fake sources using real image data (PSF and stuff)
##
cat_glob = glob('../data/stamp_catalog/cat*.fits')[1:2]
srcs, imgs, teff_catalog, us = load_imgs_and_catalog(cat_glob)
print "initialized with %d sources and %d images"%(len(srcs), len(imgs))
## pixel location of source in image 0
u_pixel = imgs[0].equa2pixel(srcs[0].u)
## generate fake sources information
gt_srcs = []
## generate a fake source (that we want to die)
t_new = np.array([5500]) # synthetic temperatures
b_new = np.array([5e-10]) # synthetic brightnesses
for s in range(1):
us_equa = imgs[0].pixel2equa(u_pixel)
src_s = celeste.PointSrcParams(u = us_equa, b = b_new[s], t = t_new[s])
#srcs.append(src_s)
# re-generate images using these source params
for img in imgs:
mimg = celeste.gen_model_image(gt_srcs, img)
img.nelec = np.random.poisson(mimg)
# check out the first image
#ll_trace, conv = celeste_em.celeste_em(srcs, imgs, maxiter=40, debug=False, verbose=1)
#compare_to_model(gt_srcs, imgs[0])
#plt.show()
##
## Initialize a single source with EM
##
np.random.seed(42)
rand = np.random.RandomState()
Niter = 1
e_samps = np.zeros((Niter, len(imgs))) # num images is fixed
ll_samps = np.zeros(Niter)
post_samps = [srcs] #
for iter_n in range(Niter):
print "number of sources: ", len(srcs)
# fresh new universe model
srcs = deathStar(srcs, imgs, rand=rand)
# keep track of universe model and image specific noise params
post_samps.append(srcs)
for n_img in range(len(imgs)):
e_samps[iter_n, n_img] = imgs[n_img].epsilon
ll_samps[iter_n] = celeste.celeste_likelihood_multi_image(srcs, imgs)
def testSplit():
np.random.seed(42)
##
## Generate some fake sources using real image data (PSF and stuff)
##
cat_glob = glob('data/stamp_catalog/cat*.fits')[1:2]
srcs, imgs, teff_catalog, us = load_imgs_and_catalog(cat_glob)
print "initialized with %d sources and %d images"%(len(srcs), len(imgs))
## pixel location of source in image 0
u_pixel = imgs[0].equa2pixel(srcs[0].u)
## generate fake sources information
t_gt = np.array([5500, 8500]) # synthetic temperatures
b_gt = np.array([5e-10, 8e-10]) # synthetic brightnesses
gt_srcs = []
for s in range(2):
us_pixel = u_pixel + 3*np.random.randn(2)
us_equa = imgs[0].pixel2equa(us_pixel)
src_s = celeste.PointSrcParams(u = us_equa, b = b_gt[s], t = t_gt[s])
gt_srcs.append(src_s)
srcs = [celeste.PointSrcParams(u=u, b=6e-10, t=6500)]
# re-generate images using these source params
for img in imgs:
mimg = celeste.gen_model_image(gt_srcs, img)
img.nelec = np.random.poisson(mimg)
# check out the first image
compare_to_model(gt_srcs, imgs[0])
plt.show()
np.random.seed(42)
rand = np.random.RandomState()
e_samps = np.zeros(100, len(imgs))) # num images is fixed
ll_samps = np.zeros(100)
iter_n = 1
while len(srcs) == 1:
print "Doing iteration", iter_n
srcs = copy.deepcopy(srcs)
srcs = splitStar(srcs, imgs, rand=rand)
post_samps.append(srcs)
for n_img in range(len(imgs)):
e_samps[iter_n, n_img] = imgs[n_img].epsilon
ll_samps[iter_n] = celeste.celeste_likelihood_multi_image(srcs, imgs)
print "After iteration %s: %s, cat len %s" % (iter_n, ll_samps[iter_n], len(srcs))
print "Location: ", [src.u for src in srcs]
print "Brightness: ", [src.b for src in srcs]
iter_n += 1
s.fluxes = None
for img in imgs:
img.epsilon = np.random.rand() * 1e3
# save ground truth values
t_gt = np.array([s.t for s in srcs])
b_gt = np.array([s.b for s in srcs])
e_gt = np.array([img.epsilon for img in imgs])
# re-generate images using these source params
for img in imgs:
mimg = gen_model_image(srcs, img)
img.nelec = np.random.poisson(mimg)
#cache initial likelihood, initial temp
ll0 = celeste_likelihood_multi_image(srcs, imgs)
print "Data generating log like: ", ll0
# re-initialize source values randomly
for s in srcs:
s.t = np.random.rand() * 9e3 + 1000
s.b = np.random.rand() * 1e-9
for img in imgs:
img.epsilon = np.random.rand() * 1e3
print "random initial ll = %2.2f" % celeste_likelihood_multi_image(
srcs, imgs)
print "========= EM for %d sources, %d images ================" % (
len(srcs), len(imgs))
s.fluxes = None
for img in imgs:
img.epsilon = np.random.rand()*1e3
# save ground truth values
t_gt = np.array([s.t for s in srcs])
b_gt = np.array([s.b for s in srcs])
e_gt = np.array([img.epsilon for img in imgs])
# re-generate images using these source params
for img in imgs:
mimg = gen_model_image(srcs, img)
img.nelec = np.random.poisson(mimg)
#cache initial likelihood, initial temp
ll0 = celeste_likelihood_multi_image(srcs, imgs)
print "Data generating log like: ", ll0
# re-initialize source values randomly
for s in srcs:
s.t = np.random.rand()*9e3 + 1000
s.b = np.random.rand()*1e-9
for img in imgs:
img.epsilon = np.random.rand()*1e3
print "random initial ll = %2.2f"%celeste_likelihood_multi_image(srcs, imgs)
print "========= EM for %d sources, %d images ================"%(len(srcs), len(imgs))
#%prun -s tottime ll_trace, conv = celeste_em(srcs, imgs, 2, debug=False, verbose=1)
ll_trace, conv = celeste_em(srcs, imgs, maxiter=40, debug=False, verbose=1)
for img in imgs:
img.epsilon = np.random.rand()*1e3
# save ground truth values
t_gt = np.array([s.t for s in srcs])
teff_catalog = t_gt
b_gt = np.array([s.b for s in srcs])
e_gt = np.array([img.epsilon for img in imgs])
# re-generate images using these source params
for img in imgs:
mimg = gen_model_image(srcs, img)
img.nelec = np.random.poisson(mimg)
#cache initial likelihood, initial temp
%lprun -m celeste ll0 = celeste_likelihood_multi_image(srcs, imgs)
print "Data generating log like: ", ll0
# re-initialize source values randomly
for s in srcs:
s.t = np.random.rand()*9e3 + 1000
s.b = np.random.rand()*1e-9
s.fluxes = None
for img in imgs:
img.epsilon = np.random.rand()*1e3
print "========= EM for %d sources, %d images ================"%(len(srcs), len(imgs))
#%prun -s tottime ll_trace, conv = celeste_em(srcs, imgs, 2, debug=False, verbose=1)
ll_trace, conv = celeste_em(srcs, imgs, maxiter=10, debug=False, verbose=1)
for s in srcs: