def knn_test(Xtrain, Ytrain, Xtest, pp, opts):
for metric in opts.metric:
tic = time()
D = pairwise_dists(Xtest, Xtrain, metric, num_procs=opts.parallel)
d_time = time() - tic
for k, w in cart_product(opts.k, opts.weights):
yield _test_knn(D, Ytrain, k, w, metric, d_time, pp, opts)
def hyper_parameter_grid_search(
data: dataUtils.Data,
split_index: int,
hyper_parameter_space: dict,
show_status: bool = False,
):
"""Perform hyperparamter grid search optimization
:param data: Data object
:param split_index: split the data into training and test set
:param hyper_parameter_space: a dictionary containing the limitis of the hyper parameter search
of the form {"hyperparamter": min, max, no. steps}
:param show_status: determine whether you want the result of each run printed to the console or not.
:return: The set of optimal hyperparameters
"""
print("Run Grid Search")
data.split_data_by_index(split_index)
search_points = dict()
for param in hyper_parameter_space.keys():
p_start = hyper_parameter_space[param][0]
p_stop = hyper_parameter_space[param][1]
p_num = hyper_parameter_space[param][2]
search_points[param] = np.linspace(start=p_start, stop=p_stop, num=p_num)
if param in ["internal_nodes", "seed"]:
search_points[param] = list(map(lambda x: int(x), search_points[param]))
parameter_combinations = cart_product(
*[search_points[param] for param in sorted(search_points.keys())]
)
min_error: float = float("inf")
optimal_hyper_parameter: dict = dict()
run_id: int = 0
for combination in parameter_combinations:
run_id += 1
hyper_parameter_sample = dict(zip(sorted(search_points.keys()), combination))
test_error = __evaluate_esn_model(data, hyper_parameter_sample)
if np.abs(test_error) < min_error:
min_error = np.abs(test_error)
optimal_hyper_parameter = hyper_parameter_sample
if show_status:
status = " %.d " % run_id
for value in combination:
status += " %.2f" % value
print(
status
+ " Error: %.4f Global Optimum: %.4f" % (np.abs(test_error), min_error)
)
print("Grid Search Completed. Min Error %f" % min_error)
pprint(optimal_hyper_parameter)
return min_error, optimal_hyper_parameter
def generate_boxes(shape, kernel):
if len(shape) != len(kernel):
raise AssertionError(
"Incompatible shape ({}) and kernel ({})".format(
shape, kernel))
ticks = [[(x, x + step) for x in range(0, axis, step)]
for axis, step in zip(shape, kernel)]
return list(cart_product(*ticks))
def test_series_group_min_max(self):
for op, level, skipna in cart_product(self.AGG_FUNCTIONS,
range(2),
[False, True]):
grouped = self.series.groupby(level=level)
aggf = lambda x: getattr(x, op)(skipna=skipna)
# skipna=True
leftside = grouped.agg(aggf)
rightside = getattr(self.series, op)(level=level, skipna=skipna)
assert_series_equal(leftside, rightside)
def factor_at_nonzero(P):
n = len(P)
Z = point_factors(P, 0)
for i in range(2, n // 2 + 1):
M = point_factors(P, i)
Z = [flatten(z) for z in cart_product(Z, M)]
for u in Z:
Q = Polynomial(u)
S = P / Q
if len(S) == i and S.coef[-1] == 0:
return Q
return P
def test_frame_group_ops(self):
self.frame.ix[1, [1, 2]] = np.nan
self.frame.ix[7, [0, 1]] = np.nan
for op, level, axis, skipna in cart_product(self.AGG_FUNCTIONS,
range(2), range(2),
[False, True]):
if axis == 0:
frame = self.frame
else:
frame = self.frame.T
grouped = frame.groupby(level=level, axis=axis)
aggf = lambda x: getattr(x, op)(skipna=skipna, axis=axis)
leftside = grouped.agg(aggf)
rightside = getattr(frame, op)(level=level, axis=axis,
skipna=skipna)
assert_frame_equal(leftside, rightside)
def removePatternNoise(self, adinputs=None, **params):
"""
This attempts to remove the pattern noise in NIRI/GNIRS data. In each
quadrant, boxes of a specified size are extracted and, for each pixel
location in the box, the median across all the boxes is determined.
The resultant median is then tiled to the size of the quadrant and
subtracted. Optionally, the median of each box can be subtracted
before performing the operation.
Based on Andy Stephens's "cleanir"
Parameters
----------
suffix: str
suffix to be added to output files
force: bool
perform operation even if standard deviation in quadrant increases?
hsigma/lsigma: float
sigma-clipping limits
pattern_x_size: int
size of pattern "box" in x direction
pattern_y_size: int
size of pattern "box" in y direction
subtract_background: bool
remove median of each "box" before calculating pattern noise?
"""
log = self.log
log.debug(gt.log_message("primitive", self.myself(), "starting"))
timestamp_key = self.timestamp_keys[self.myself()]
hsigma, lsigma = params["hsigma"], params["lsigma"]
pxsize, pysize = params["pattern_x_size"], params["pattern_y_size"]
bgsub = params["subtract_background"]
force = params["force"]
stack_function = NDStacker(combine='median', reject='sigclip',
hsigma=hsigma, lsigma=lsigma)
sigclip = partial(sigma_clip, sigma_lower=lsigma, sigma_upper=hsigma)
zeros = None # will remain unchanged if not subtract_background
for ad in adinputs:
if ad.phu.get(timestamp_key):
log.warning("No changes will be made to {}, since it has "
"already been processed by removePatternNoise".
format(ad.filename))
continue
for ext in ad:
qysize, qxsize = [size // 2 for size in ext.data.shape]
yticks = [(y, y + pysize) for y in range(0, qysize, pysize)]
xticks = [(x, x + pxsize) for x in range(0, qxsize, pxsize)]
for ystart in (0, qysize):
for xstart in (0, qxsize):
quad = ext.nddata[ystart:ystart + qysize, xstart:xstart + qxsize]
sigma_in = sigclip(np.ma.masked_array(quad.data, quad.mask)).std()
# print sigma_in
blocks = [quad[tuple(slice(start, end)
for (start, end) in coords)]
for coords in cart_product(yticks, xticks)]
if bgsub:
# If all pixels are masked in a box, we'll get no
# result from the mean. Suppress warning.
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UserWarning)
zeros = np.nan_to_num([-np.ma.masked_array(block.data, block.mask).mean()
for block in blocks])
out = stack_function(blocks, zero=zeros).data
out_quad = (quad.data + np.mean(out) -
np.tile(out, (len(yticks), len(xticks))))
sigma_out = sigclip(np.ma.masked_array(out_quad, quad.mask)).std()
if sigma_out > sigma_in:
qstr = (f"{ad.filename} extension {ext.id} "
f"quadrant ({xstart},{ystart})")
if force:
log.stdinfo("Forcing cleaning on " + qstr)
else:
log.stdinfo("No improvement for "+qstr)
continue
ext.data[ystart:ystart + qysize, xstart:xstart + qxsize] = out_quad
# Timestamp and update filename
gt.mark_history(ad, primname=self.myself(), keyword=timestamp_key)
ad.update_filename(suffix=params["suffix"], strip=True)
return adinputs
def init_default_extensions(self, num_ext=12, binning=1, overscan=True,
read_speed="slow", gain_setting="low"):
if num_ext != 12:
raise NotImplementedError("Only tested for full array ROI")
if binning not in (1, 2, 4):
raise ValueError("Binning must be 1, 2, or 4")
del self[:]
shape = ((4224 if self.phu['DETID'].startswith('BI') else 4608) // binning,
512 // binning + (BIAS_WIDTH if overscan else 0))
# If the overscan is present, assume it's raw data
dtype = np.uint16 if overscan else np.float32
pixel_scale = lookup.gmosPixelScales[self.instrument(),
self.phu['DETTYPE']] * binning
north = self.instrument() == 'GMOS-N'
crpix1 = 3132.69 if north else 3133.5
crpix1 = (crpix1 - 0.5) / binning + 0.5
crpix2_list = CRPIX2N if north else CRPIX2S
chip_gap = 67. if north else 61.
extra_keywords = {'CRVAL1': self.phu['RA'], 'CRVAL2': self.phu['DEC'],
'CTYPE1': 'RA---TAN', 'CTYPE2': 'DEC--TAN',
'CCDSUM': '{} {}'.format(binning, binning)}
self.phu['NAMPS'] = num_ext
for i in range(num_ext):
ccd = i // 4
crpix2 = (crpix2_list[ccd] - 0.5) / binning + 0.5
detx1 = i * 512
detx2 = detx1 + 512
detsec = '[{}:{},1:4224]'.format(detx1 + 1, detx2)
datx1 = BIAS_WIDTH if (overscan and i % 2 == 1) else 0
datx2 = datx1 + 512 // binning
datasec = '[{}:{},1:{}]'.format(datx1 + 1, datx2, shape[0])
if overscan:
biasx1 = BIAS_WIDTH - datx1
biassec = '[{}:{},1:{}]'.format(biasx1 + 1, biasx1 + BIAS_WIDTH, shape[0])
extra_keywords[self._keyword_for('overscan_section')] = biassec
arrx1 = detx1 % 2048
arrx2 = arrx1 + 512
arraysec = '[{}:{},1:{}]'.format(arrx1 + 1, arrx2, shape[0])
extra_keywords.update({self._keyword_for('detector_section'): detsec,
self._keyword_for('data_section'): datasec,
self._keyword_for('array_section'): arraysec,
'CRPIX1': crpix1 + datx1, 'CRPIX2': crpix2})
crpix1 -= (datx2 - datx1)
# This isn't entirely right but it'll do
if i % 4 == 3:
crpix1 -= chip_gap / binning
self.add_extension(shape=shape, pixel_scale=pixel_scale,
dtype=dtype, extra_keywords=extra_keywords)
# GAIN and READNOISE
# not the correct values, but makes the descriptors work
self.phu['AMPINTEG'] = 10000 if read_speed == "slow" else 1000
self.hdr['GAIN'] = 1 if gain_setting == "low" else 5
ccdnames = self.phu['DETID'].split(",")
if len(ccdnames) > 1:
if ccdnames[0].startswith("e2v"):
ccdnames[1] = "e2v " + ccdnames[1]
ccdnames[2] = "e2v " + ccdnames[2]
else:
ccdnames = ["EEV"+x for x in self.phu['DETID'].split("EEV")[1:]]
if num_ext == 12:
amps = ("1", "2", "3", "4")
else:
amps = ("left", "right")
for i, (ccd, amp) in enumerate(cart_product(ccdnames, amps)):
self[i].hdr['AMPNAME'] = f"{ccd}, {amp}"
def removePatternNoise(self, adinputs=None, **params):
"""
This attempts to remove the pattern noise in NIRI/GNIRS data. In each
quadrant, boxes of a specified size are extracted and, for each pixel
location in the box, the median across all the boxes is determined.
The resultant median is then tiled to the size of the quadrant and
subtracted. Optionally, the median of each box can be subtracted
before performing the operation.
Based on Andy Stephens's "cleanir"
Parameters
----------
suffix: str
suffix to be added to output files
force: bool
perform operation even if standard deviation in quadrant increases?
hsigma/lsigma: float
sigma-clipping limits
pattern_x_size: int
size of pattern "box" in x direction
pattern_y_size: int
size of pattern "box" in y direction
subtract_background: bool
remove median of each "box" before calculating pattern noise?
"""
log = self.log
log.debug(gt.log_message("primitive", self.myself(), "starting"))
timestamp_key = self.timestamp_keys[self.myself()]
hsigma, lsigma = params["hsigma"], params["lsigma"]
pxsize, pysize = params["pattern_x_size"], params["pattern_y_size"]
bgsub = params["subtract_background"]
force = params["force"]
stack_function = NDStacker(combine='median', reject='sigclip',
hsigma=hsigma, lsigma=lsigma)
sigclip = partial(sigma_clip, sigma_lower=lsigma, sigma_upper=hsigma)
zeros = None # will remain unchanged if not subtract_background
for ad in adinputs:
if ad.phu.get(timestamp_key):
log.warning("No changes will be made to {}, since it has "
"already been processed by removePatternNoise".
format(ad.filename))
continue
for ext in ad:
qysize, qxsize = [size // 2 for size in ext.data.shape]
yticks = [(y, y + pysize) for y in range(0, qysize, pysize)]
xticks = [(x, x + pxsize) for x in range(0, qxsize, pxsize)]
for ystart in (0, qysize):
for xstart in (0, qxsize):
quad = ext.nddata[ystart:ystart + qysize, xstart:xstart + qxsize]
sigma_in = sigclip(np.ma.masked_array(quad.data, quad.mask)).std()
# print sigma_in
blocks = [quad[tuple(slice(start, end)
for (start, end) in coords)]
for coords in cart_product(yticks, xticks)]
if bgsub:
# If all pixels are masked in a box, we'll get no
# result from the mean. Suppress warning.
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=UserWarning)
zeros = np.nan_to_num([-np.ma.masked_array(block.data, block.mask).mean()
for block in blocks])
out = stack_function(blocks, zero=zeros).data
out_quad = (quad.data + np.mean(out) -
np.tile(out, (len(yticks), len(xticks))))
sigma_out = sigclip(np.ma.masked_array(out_quad, quad.mask)).std()
if sigma_out > sigma_in:
qstr = "{}:{} quadrant ({},{})".format(ad.filename,
ext.hdr['EXTVER'], xstart, ystart)
if force:
log.stdinfo("Forcing cleaning on " + qstr)
else:
log.stdinfo("No improvement for "+qstr)
continue
ext.data[ystart:ystart + qysize, xstart:xstart + qxsize] = out_quad
# Timestamp and update filename
gt.mark_history(ad, primname=self.myself(), keyword=timestamp_key)
ad.update_filename(suffix=params["suffix"], strip=True)
return adinputs