def __call__(self, logits, ys):
forward = solution.SigmoidCrossEntropyForward()
backward = solution.SigmoidCrossEntropyBackward()
ctx = dict()
average_loss = forward(logits, ys, ctx=ctx)
dlogits = backward(ctx, 1.0)
for d in range(self.batch_size):
real_dlogits = np.zeros_like(logits)
real_dlogits[d] = self.delta
moved_logits = logits + real_dlogits
moved_average_loss = forward(moved_logits, ys)
assert np.all(
np.absolute((moved_average_loss - average_loss) / self.delta -
dlogits[d]) < 1e-5
), 'Value mismatch in dimension {}.'.format(d)
def __init__(self):
self._forward = solution.SigmoidCrossEntropyForward()
self._backward = solution.SigmoidCrossEntropyBackward()
def __call__(self):
forward = solution.SigmoidCrossEntropyForward()
self.positive_monotonically_decreasing(forward)
self.negative_monotonically_increasing(forward)