|
1
|
1
|
import warnings
|
|
2
|
|
|
|
3
|
1
|
import numpy as np
|
|
4
|
|
|
|
5
|
1
|
from nengo.solvers import Solver, LstsqL2
|
|
6
|
1
|
from nengo.version import version_info
|
|
7
|
|
|
|
8
|
|
if version_info < (2, 5, 0): # pragma: no cover
|
|
9
|
|
Temporal = None # not supported (requires nengo PR #1313)
|
|
10
|
|
else:
|
|
11
|
1
|
from nengolib.temporal import Temporal
|
|
12
|
|
|
|
13
|
1
|
__all__ = ['Temporal']
|
|
14
|
|
|
|
15
|
|
|
|
16
|
1
|
class BiasedSolver(Solver):
|
|
17
|
|
"""Wraps a solver with a bias neuron, and extracts its weights.
|
|
18
|
|
|
|
19
|
|
This is setup correctly by nengolib.Connection; not to be used directly.
|
|
20
|
|
"""
|
|
21
|
|
|
|
22
|
1
|
def __init__(self, solver=LstsqL2()):
|
|
23
|
1
|
self.solver = solver
|
|
24
|
1
|
self.bias = None
|
|
25
|
1
|
try:
|
|
26
|
|
# parent class changed in Nengo 2.1.1
|
|
27
|
|
# need to do this because self.weights is read-only
|
|
28
|
1
|
super(BiasedSolver, self).__init__(weights=solver.weights)
|
|
29
|
|
except TypeError: # pragma: no cover
|
|
30
|
|
super(BiasedSolver, self).__init__()
|
|
31
|
|
self.weights = solver.weights
|
|
32
|
|
|
|
33
|
1
|
def __call__(self, A, Y, __hack__=None, **kwargs):
|
|
34
|
1
|
assert __hack__ is None
|
|
35
|
|
# __hack__ is necessary prior to nengo PR #1359 (<2.6.1)
|
|
36
|
|
# and following nengo PR #1507 (>2.8.0)
|
|
37
|
|
|
|
38
|
1
|
if self.bias is not None:
|
|
39
|
|
# this is okay if due to multiple builds of the same network (#99)
|
|
40
|
1
|
warnings.warn("%s called twice; ensure not being shared between "
|
|
41
|
|
"multiple connections" % type(self).__name__,
|
|
42
|
|
UserWarning)
|
|
43
|
1
|
scale = A.max() # to make regularization consistent
|
|
44
|
1
|
AB = np.empty((A.shape[0], A.shape[1] + 1))
|
|
45
|
1
|
AB[:, :-1] = A
|
|
46
|
1
|
AB[:, -1] = scale
|
|
47
|
1
|
XB, solver_info = self.solver.__call__(AB, Y, **kwargs)
|
|
48
|
1
|
solver_info['bias'] = self.bias = XB[-1, :] * scale
|
|
49
|
1
|
return XB[:-1, :], solver_info
|
|
50
|
|
|
|
51
|
1
|
def bias_function(self, size):
|
|
52
|
|
"""Returns the function for the pre-synaptic bias node."""
|
|
53
|
1
|
return lambda _: np.zeros(size) if self.bias is None else self.bias
|
arvoelke