def ImageLayerFromInputArray(input_, backend):
"""Create the initial image layer from some input.
:param input_: Input data. If array, values should lie in the range [0, 1].
:type input_: PIL.Image or 2D ndarray of float
:returns: Image layer data.
:rtype: 2D ndarray of float
"""
if isinstance(input_, Image.Image):
input_ = input_.convert('L')
input_ = ImageToArray(input_, transpose = True)
input_ = input_.astype(np.float)
# Map from [0, 255] to [0, 1]
input_ /= 255
# Note: we could have scaled pixels to [-1, 1]. This would result in a
# doubling of the dynamic range of the S1 response, since each S1 activation
# is of the form:
# y = XW
# where X and W are the input and weight vector, respectively. The rescaled
# version of X (call it X') is given by:
# X' = 2X - 1
# so the activation is given by
# y' = X'W = (2X - 1)W = 2XW - \sum w_i = 2XW
# since W is a mean-zero Gabor filter. (This ignores retinal processing, and
# nonlinearities caused by normalization). The scaling of S1 response seems
# unlikely to cause a significant change in the network output.
return backend.PrepareArray(input_)
def ImageLayerFromInputArray(input_, backend):
"""Create the initial image layer from some input.
:param input_: Input data. If array, values should lie in the range [0, 1].
:type input_: PIL.Image or 2D ndarray of float
:returns: Image layer data.
:rtype: 2D ndarray of float
"""
if isinstance(input_, Image.Image):
input_ = input_.convert('L')
input_ = ImageToArray(input_, transpose=True)
input_ = input_.astype(np.float)
# Map from [0, 255] to [0, 1]
input_ /= 255
# Note: we could have scaled pixels to [-1, 1]. This would result in a
# doubling of the dynamic range of the S1 response, since each S1 activation
# is of the form:
# y = XW
# where X and W are the input and weight vector, respectively. The rescaled
# version of X (call it X') is given by:
# X' = 2X - 1
# so the activation is given by
# y' = X'W = (2X - 1)W = 2XW - \sum w_i = 2XW
# since W is a mean-zero Gabor filter. (This ignores retinal processing, and
# nonlinearities caused by normalization). The scaling of S1 response seems
# unlikely to cause a significant change in the network output.
return backend.PrepareArray(input_)
def ImageLayerFromInputArray(input_, backend):
"""Create the initial image layer from some input.
input_ -- Image or (2-D) array of input data. If array, values should lie in
the range [0, 1].
RETURNS (2-D) array containing image layer data
"""
if isinstance(input_, Image.Image):
input_ = input_.convert('L')
input_ = ImageToArray(input_, transpose = True)
input_ = input_.astype(np.float)
# Map from [0, 255] to [0, 1]
input_ /= 255
return backend.PrepareArray(input_)
def ImageLayerFromInputArray(input_, backend):
"""Create the initial image layer from some input.
:param input_: Input data. If array, values should lie in the range [0, 1].
:type input_: PIL.Image or 2D ndarray of float
:returns: Image layer data.
:rtype: 2D ndarray of float
"""
if isinstance(input_, Image.Image):
input_ = input_.convert('L')
input_ = ImageToArray(input_, transpose=True)
input_ = input_.astype(np.float)
# Map from [0, 255] to [0, 1]
input_ /= 255
return backend.PrepareArray(input_)
def ImageLayerFromInputArray(input_, backend):
"""Create the initial image layer from some input.
:param input_: Input data. If array, values should lie in the range [0, 1].
:type input_: PIL.Image or 2D ndarray of float
:returns: Image layer data.
:rtype: 2D ndarray of float
"""
if isinstance(input_, Image.Image):
input_ = input_.convert('L')
input_ = ImageToArray(input_, transpose = True)
input_ = input_.astype(np.float)
# Map from [0, 255] to [0, 1]
input_ /= 255
return backend.PrepareArray(input_)
