def Dgrade(nside_in, nside_out):
"""Keras layer performing a downgrade of input maps
Parameters
----------
nside_in : integer
Nside parameter for the input maps.
Must be a valid healpix Nside value
nside_out: integer
Nside parameter for the output maps.
Must be a valid healpix Nside value
"""
file_in = os.path.join(
os.path.dirname(__file__),
"..",
"ancillary_files",
"dgrade_from{}_to{}.npy".format(nside_in, nside_out),
)
try:
pixel_indices = np.load(file_in)
except:
pixel_indices = nnh.dgrade(nside_in, nside_out)
def f(x):
y = OrderMap(pixel_indices)(x)
pool_size = int((nside_in / nside_out)**2.0)
y = tf.keras.layers.AveragePooling1D(pool_size=pool_size)(y)
return y
return f
def Pooling(nside_in, nside_out, layer1D, *args, **kwargs):
"""Keras layer performing a downgrade+custom pooling of input maps
Args:
* nside_in (integer): ``NSIDE`` parameter for the input maps.
* nside_out (integer): ``NSIDE`` parameter for the output maps.
* layer1D (layer object): a 1-D layer operation, like
:code:`keras.layers.MaxPooling1D`
* args (any): Positional arguments to be passed to :code:`layer1D`
* kwargs: keyword arguments to be passed to
:code:`layer1D`. The keyword :code:`pool_size` should not be
included, as it is handled automatically.
"""
file_in = os.path.join(
os.path.dirname(__file__),
"..",
"ancillary_files",
"dgrade_from{}_to{}.npy".format(nside_in, nside_out),
)
try:
pixel_indices = np.load(file_in)
except:
pixel_indices = nnh.dgrade(nside_in, nside_out)
def f(x):
y = OrderMap(pixel_indices)(x)
pool_size = int((nside_in / nside_out)**2.0)
kwargs["pool_size"] = pool_size
y = layer1D(*args, **kwargs)(y)
return y
return f
def DegradeAndConvNeighbours(nside_in,
nside_out,
filters,
use_bias=False,
trainable=True):
"""Keras layer performing a downgrading and convolution of input maps.
Args:
* nside_in (integer): ``NSIDE`` parameter for the input maps.
* nside_out (integer): ``NSIDE`` parameter for the output maps.
* filters (integer): Number of filters to use in the
convolution
* use_bias (bool): Whether the layer uses a bias vector or
not. Default is ``False``.
* trainable (bool): Wheter this is a trainable layer or
not. Default is ``True``.
"""
file_in = os.path.join(
os.path.dirname(__file__),
"..",
"ancillary_files",
"dgrade_from{}_to{}.npy".format(nside_in, nside_out),
)
try:
pixel_indices = np.load(file_in)
except:
pixel_indices = nnh.dgrade(nside_in, nside_out)
def f(x):
y = OrderMap(pixel_indices)(x)
kernel_size = int((nside_in / nside_out)**2.0)
y = Conv1D(
filters,
kernel_size=kernel_size,
strides=kernel_size,
use_bias=use_bias,
trainable=trainable,
kernel_initializer="random_uniform",
)(y)
return y
return f