def test_centroid(self):
"""
Test that the centroiding operation works as expected.
"""
# Start by building a sample test image.
nx, ny = 150, 200
X, Y = np.meshgrid(range(nx), range(ny))
centers = np.array([[50.0, 45.0], [75.6, 150.1]])
widths = [10, 4]
amps = [5, 2]
img = np.zeros((ny, nx))
for i, p in enumerate(centers):
R2 = (X - p[0])**2 + (Y - p[1])**2
img += amps[i] * np.exp(-0.5 * R2 / widths[i]**2) / widths[i]
# Then try a centered scene.
nsx, nsy = 45, 57
SX, SY = np.meshgrid(range(nsx), range(nsy))
scene = np.exp(-0.5 * ((SX - (nsx - 1) / 2) ** 2 \
+ (SY - (nsy - 1) / 2) ** 2))
# Run the centroid.
size = 24
coords, data, mask = utils.centroid_image(img, size, scene=scene)
# Make sure that it gets the right coordinates.
truth = centers[np.argmax(amps)]
assert np.all(coords[::-1] == truth)
def test_centroid(self):
"""
Test that the centroiding operation works as expected.
"""
# Start by building a sample test image.
nx, ny = 150, 200
X, Y = np.meshgrid(range(nx), range(ny))
centers = np.array([[50.0, 45.0], [75.6, 150.1]])
widths = [10, 4]
amps = [5, 2]
img = np.zeros((ny, nx))
for i, p in enumerate(centers):
R2 = (X - p[0]) ** 2 + (Y - p[1]) ** 2
img += amps[i] * np.exp(-0.5 * R2 / widths[i] ** 2) / widths[i]
# Then try a centered scene.
nsx, nsy = 45, 57
SX, SY = np.meshgrid(range(nsx), range(nsy))
scene = np.exp(-0.5 * ((SX - (nsx - 1) / 2) ** 2 + (SY - (nsy - 1) / 2) ** 2))
# Run the centroid.
size = 24
coords, data, mask = utils.centroid_image(img, size, scene=scene)
# Make sure that it gets the right coordinates.
truth = centers[np.argmax(amps)]
assert np.all(coords[::-1] == truth)
def do_update(self, fn, alpha, maskfn=None, maskhdu=0, median=True,
nn=False, hack=True):
"""
Do a single stochastic gradient update using the image in a
given file and learning rate.
## Arguments
* `fn` (str): The filename of the image to be used.
* `alpha` (float): The learning rate.
## Keyword Arguments
* `maskfn` (str): Path to the mask file.
* `maskhdu` (int): The HDU number for the mask.
* `median` (bool): Subtract the median of the scene?
* `nn` (bool): Project onto the non-negative plane?
## Returns
* `data` (numpy.ndarray): The centered and cropped data image used
for this update.
"""
image = utils.load_image(fn, hdu=self.hdu)
if maskfn is not None:
mask = utils.load_image(maskfn, hdu=maskhdu)
if self.invert:
inds = np.isnan(mask) + np.isinf(mask)
mask[~inds] = 1.0 / mask[~inds]
mask[inds] = 0.0
if self.square:
mask = mask ** 2
else:
mask = np.ones_like(image)
# Center the data.
if self.centers is None:
data = utils.trim_image(image, self.size)
mask = utils.trim_image(mask, self.size)
else:
result = utils.centroid_image(image, self.size,
coords=self.centers[fn], mask=mask)
data = result[1]
mask = result[2]
# Deal with NaNs and infinities.
if maskfn is None:
mask *= ~(np.isnan(data) + np.isinf(data))
data[mask == 0.0] = 0.0
assert np.all(~(np.isnan(data) + np.isinf(data))), \
"Yer data's got some unmasked NaNs or infs, dude."
# Add the DC offset.
data += self.dc
# Do the inference.
self.old_scene = np.array(self.scene)
self.psf, self.sky = self.infer_psf(data, mask)
if hack:
logging.info("Hacking.")
X, Y = np.meshgrid(np.arange(-self.psf_hw, self.psf_hw + 1),
np.arange(-self.psf_hw, self.psf_hw + 1))
R = np.sqrt(X ** 2 + Y ** 2)
outer = self.psf[R > 0.9 * self.psf_hw]
inds = outer > 0
if np.any(inds):
self.psf -= np.sum(inds) / float(outer.size) \
* np.median(outer[inds])
print "sky:", self.sky
self.dlds = self.get_dlds(data, mask)
# self.old_scene = self.scene + alpha * self.dlds
self.scene += alpha * self.dlds
# WTF?!?
gc.collect()
# Apply some serious HACKS!
if median:
self.scene -= np.median(self.scene)
if nn:
self.scene[self.scene < 0] = 0.0
return data