def test(self):
pred_batch = share(np.reshape(np.array([0, 0.2, 0.8, 0, 0.6, 0.4]), (2,3)))
tg_batch = share(np.reshape(np.array([0, 0, 1, 0, 0, 1]), (2,3)))
a = T.argmax(pred_batch, axis=1)
b = T.argmax(tg_batch, axis=1)
weights = 1 + 10 * (self.volumes[a] / self.volumes[b]) * (self.n/self.m)
return -T.mean(weights * T.log(T.sum(pred_batch * tg_batch, axis=1)))
def __init__(self, volumes):
CostFunction.__init__(self)
self.m = share(np.max(volumes))
self.n = share(np.min(volumes))
volumes = np.concatenate([np.array([0]), volumes])
self.volumes = share(volumes)
def init_parameters(self, w_shape, b_shape):
w_bound = self.compute_bound_parameters_virtual()
# initialize weights with random weights
self.w = share(np.asarray(
np.random.uniform(low=-w_bound, high=w_bound, size=w_shape),
dtype=theano.config.floatX), "w")
# the bias is a 1D tensor -- one bias per output feature map
b_values = 0.1 + np.zeros(b_shape, dtype=theano.config.floatX) # Slightly positive for RELU units
self.b = share(b_values, "b")
self.update_params()
def outputs_shared(self):
if not self.shared_outputs_created:
self._outputs_shared = share(self.outputs)
self.shared_outputs_created = True
else:
self._outputs_shared.set_value(self.outputs, borrow=True)
return self._outputs_shared
def compute_updates(self, params, grads):
updates = OrderedDict()
for param_i, grad_i in zip(params, grads):
diff = share(np.zeros(param_i.get_value().shape, dtype=theano.config.floatX), "diff")
update_diff = self.momentum * diff - self.learning_rate * grad_i
updates[param_i] = param_i + update_diff
updates[diff] = update_diff
return updates
def __init__(self, vec):
LayerBlock.__init__(self)
self.vec = share(vec)
def __init__(self, learning_rate, momentum):
LearningUpdate.__init__(self)
self.learning_rate = share(learning_rate, "learning_rate")
self.momentum = share(momentum, "momentum")
if momentum < 0 or momentum > 1:
raise Exception("Momentum value should be between 0 and 1.")
def __init__(self, learning_rate):
LearningUpdate.__init__(self)
self.learning_rate = share(learning_rate, "learning_rate")
