def _sample_labels(self, x, s):
if self._bias:
x = x[:, 1]
yprob = 0.8 * sigmoid(0.6 *
(x + 3)) * sigmoid(-5 * (x - 3)) + sigmoid(x - 5)
return np.expand_dims(np.random.binomial(1, yprob), axis=1)
def calc_effort(self, inputs):
Z = np.concatenate((np.array(inputs[:self.n_input]), self.hidden_out))
preact = Z.dot(self.W) + self.b
f, i, o, candidate = np.split(sigmoid(preact), 4)
self.memory *= f
self.memory += i * candidate
self.hidden_out = o * sigmoid(self.memory)
output = sigmoid(self.hidden_out.dot(self.Wo) + self.bo)
return float(output)
def calc_effort(self, inputs):
"""
Effort is calculated by forward-propagating
the inputs through the neural network
"""
stimuli = np.array(inputs)
for w, b in zip(self.weights, self.biases):
z = stimuli.dot(w) + b
stimuli = sigmoid(z)
return float(stimuli)
def forward_pass(self, inputs):
outputs = sigmoid(
np.dot(self._weights[:, :-1], inputs) + self._weights[:, -1])
sigmoid_grad = outputs * (1 - outputs)
self._grad_inputs = sigmoid_grad[
np.newaxis, :].T * self._weights[:, :-1]
self._grad_weights[:, :-1] = sigmoid_grad[np.newaxis, :].T * (
inputs[np.newaxis, :] * np.ones(self._grad_inputs.shape))
self._grad_weights[:, -1] = np.ones(self._num_neurons) * sigmoid_grad
return outputs
def _probability(self, features):
return sigmoid(np.matmul(self.feature_map(features), self._theta)).reshape(-1, 1)
def evaluate(self):
return util.sigmoid(statistics.mean([self.__do_nothing(), self.__rebalance()]))
def calc_effort(self, inputs):
Z = np.concatenate((np.array(inputs[:self.n_input]), self.hidden_out))
self.hidden_out = sigmoid(Z.dot(self.W) + self.b)
output = sigmoid(self.hidden_out.dot(self.Wo) + self.bo)
return float(output)