class InvSigmoid(nn.Module):
"""Pointwise :math:`s^{-1}(x) = \log(x / (1-x))`."""
def __init__(self, a=None):
"""Initialize a new instance.
Examples
--------
>>> invsig = InvSigmoid()
>>> xvals = [1 / (1 + np.exp(-1)), 1 / (1 + np.exp(-3))]
>>> x = autograd.Variable(torch.Tensor(xvals))
>>> invsig(x) # should be [1, 3]
Variable containing:
1.0000
3.0000
[torch.FloatTensor of size 2]
"""
super(InvSigmoid, self).__init__()
def forward(self, x):
return torch.log(x / (1 - x))
if __name__ == '__main__':
from dipp.util.testutils import run_doctests
from torch import autograd
import numpy as np
extraglobs = {'np': np, 'autograd': autograd}
run_doctests(extraglobs=extraglobs)
raise ValueError('`init_c` must be provided for `c=None`')
self.c = nn.Parameter(torch.Tensor([init_c]))
elif isinstance(c, nn.Parameter):
self.c = c
else:
assert c > 0
self.c = float(c)
# Sharing the `a` and `c` parameters with the constituting modules
self.inv_sigmoid = InvSigmoid()
self.local_sim_ax0 = HaarPSISimilarityMap(0, self.a, self.c)
self.local_sim_ax1 = HaarPSISimilarityMap(1, self.a, self.c)
self.wmap_ax0 = HaarPSIWeightMap(0)
self.wmap_ax1 = HaarPSIWeightMap(1)
def forward(self, x, y):
wmap_ax0 = self.wmap_ax0(x, y)
wmap_ax1 = self.wmap_ax1(x, y)
numer = torch.sum(
self.local_sim_ax0(x, y) * wmap_ax0 +
self.local_sim_ax1(x, y) * wmap_ax1)
denom = torch.sum(wmap_ax0 + wmap_ax1)
return (self.inv_sigmoid(numer / denom) / self.a) ** 2
if __name__ == '__main__':
from dipp.util.testutils import run_doctests
run_doctests()
num_digits = int(np.ceil(np.log10(max(len(modules) - 1, 1))))
idx_fmt = '{{:<{}}}'.format(num_digits)
for i, entry in summary.items():
if i == 0:
print('=== Summary of {} ==='.format(entry['name']))
print('Input shapes :', [shape_str(s) for s in entry['shapes_in']])
print('Output shapes:',
[shape_str(s) for s in entry['shapes_out']])
print('# of params :', entry['num_params'])
else:
print()
if verbose:
print(idx_fmt.format(i - 1) + ': ' + entry['repr'])
else:
print(idx_fmt.format(i - 1) + ': ' + entry['name'])
print('Input shapes :', [shape_str(s) for s in entry['shapes_in']])
print('Output shapes:',
[shape_str(s) for s in entry['shapes_out']])
print('# of params :', entry['num_params'])
if return_summary:
return summary
if __name__ == '__main__':
from dipp.util.testutils import run_doctests
import torch
run_doctests(extraglobs={'torch': torch, 'nn': torch.nn})