def _derivative(self, feed, input, target):
logits = _softmax(input.eval(feed))
forward_gradient = self.back(target, feed)
local_gradient = []
for i in range(input.shape[0]):
local_logits = logits[i].reshape((-1, 1))
jacob = np.eye(
local_logits.shape[0]) - np.matmul(local_logits, (0 * local_logits + 1).T)
local_gradient.append(
np.matmul(jacob, forward_gradient[i].reshape((-1, 1))).T)
local_gradient = np.concatenate(local_gradient, 0)
return local_gradient
def _eval(self, feed):
result = _softmax(self.input_list[0].eval(feed))
return result