def __call__(self, weights, xs, dlogits):
forward = solution.LinearLayerForward()
backward = solution.LinearLayerBackward()
ctx = dict()
logits = forward(weights, xs, ctx=ctx)
dw = backward(ctx, dlogits)
for d in range(self.num_features):
real_dw = np.zeros_like(weights)
real_dw[d] = self.delta
moved_weights = weights + real_dw
moved_logits = forward(moved_weights, xs)
assert np.all(
np.absolute(
(moved_logits - logits).dot(dlogits) / self.delta -
dw[d]) < 1e-5), 'Value mismatch in dimension {}.'.format(d)
def __call__(self, weights, xs):
forward = solution.LinearLayerForward()
alpha_xs = np.random.normal(size=(self.batch_size, self.num_features))
beta_xs = np.random.normal(size=(self.batch_size, self.num_features))
kalpha = np.random.normal() * 50
kbeta = np.random.normal() * 50
assert np.all(
np.absolute(
forward(weights, kalpha * alpha_xs + kbeta * beta_xs) -
(kalpha * forward(weights, alpha_xs) +
kbeta * forward(weights, beta_xs))) < 1e-10
), 'The linear layer should be linear with the input data.'
alpha_weights = np.random.normal(size=(self.num_features, ))
beta_weights = np.random.normal(size=(self.num_features, ))
kalpha = np.random.normal() * 50
kbeta = np.random.normal() * 50
assert np.all(
np.absolute(
forward(kalpha * alpha_weights + kbeta * beta_weights, xs) -
(kalpha * forward(alpha_weights, xs) +
kbeta * forward(beta_weights, xs))) < 1e-10
), 'The linear layer should be linear with the weights.'
def __init__(self, num_features, init_weight_scale=1.0):
self.weights = np.random.normal(init_weight_scale,
size=(num_features, ))
self._forward = solution.LinearLayerForward()
self._backward = solution.LinearLayerBackward()
self._update = solution.LinearLayerUpdate()