def __init__(self,
learning_rate,
decay,
momentum,
epsilon=1e-10,
use_locking=False,
centered=False,
name='RMSProp'):
super(RMSPropOptimizer, self).__init__(use_locking, name)
self.updater = updaters.RMSPropUpdater(learning_rate, decay, epsilon)
def _create_graph(self):
self.x = Tensor(shape=[None, self.img_channels, self.img_height, self.img_width]).Variable()
self.y_r = Tensor(shape=[None], name='Yr').Variable()
# As implemented in A3C paper
self.n1 = ops.Relu(ops.Conv2D([self.x] + self.weight_bias(), kernel_size=8, stride=4, num_output=16))
self.n2 = ops.Relu(ops.Conv2D([self.n1] + self.weight_bias(), kernel_size=4, stride=2, num_output=32))
self.action_index = Tensor(shape=[None, self.num_actions]).Variable()
self.d1 = ops.Relu(ops.InnerProduct([self.n2] + self.weight_bias(), num_output=256))
self.logits_v = ops.InnerProduct([self.d1] + self.weight_bias(), num_output=1)
self.cost_v = ops.L2Loss([self.y_r, self.logits_v])
self.logits_p = ops.InnerProduct([self.d1] + self.weight_bias(), num_output=self.num_actions)
if Config.USE_LOG_SOFTMAX: raise NotImplementedError()
else:
self.softmax_p = ops.Softmax(self.logits_p)
self.selected_action_prob = ops.Sum(self.softmax_p * self.action_index, axis=1)
self.cost_p_1 = ops.Log(ops.Clip(self.selected_action_prob, self.log_epsilon, None)) * \
(self.y_r - ops.StopGradient(self.logits_v))
self.cost_p_2 = ops.Sum(ops.Log(ops.Clip(self.softmax_p, self.log_epsilon, None)) *
self.softmax_p, axis=1) * (-self.beta)
self.cost_p_1_agg = ops.Sum(self.cost_p_1)
self.cost_p_2_agg = ops.Sum(self.cost_p_2)
self.cost_p = -(self.cost_p_1_agg + self.cost_p_2_agg)
self.cost_all = self.cost_p + self.cost_v
if Config.DUAL_RMSPROP: raise NotImplementedError()
else:
if Config.USE_GRAD_CLIP:
self.opt = updaters.RMSPropUpdater(decay=Config.RMSPROP_DECAY,
eps=Config.RMSPROP_EPSILON,
clip_gradient=Config.GRAD_CLIP_NORM)
else:
self.opt = updaters.RMSPropUpdater(decay=Config.RMSPROP_DECAY,
eps=Config.RMSPROP_EPSILON)
grads = T.grad(self.cost_all, self.network_params)
for p, g in zip(self.network_params, grads):
self.opt.append((p, g), lr_mult=1.0)
def __init__(self, prototxt):
super(RMSPropSolver, self).__init__(prototxt=prototxt)
self._updater = updaters.RMSPropUpdater(**self._update_param)
# generates update targets
for layer, blobs in self._net.params.items(): self._lr_blobs.extend(blobs)
for idx, blob in enumerate(self._lr_blobs):
if self._net._lr_mults[idx] > 0:
if blob.diff is None: continue
self._updater.append((blob.data, blob.diff),
self._net._lr_mults[idx], self._net._decay_mults[idx])
self.train = self._net.function
self.tests = [test_net.function for test_net in self._test_nets]
self.update = function(updater=self._updater)
def __init__(
self,
learning_rate,
decay=0.9,
momentum=0.0,
epsilon=1e-10,
use_locking=False,
centered=False,
name='RMSProp',
):
super(RMSPropOptimizer, self).__init__(use_locking, name)
if momentum > 0.:
self.updater = _updaters.AdamUpdater(
learning_rate, momentum, decay, epsilon)
else:
self.updater = _updaters.RMSPropUpdater(
learning_rate, decay, epsilon)
self._set_dynamic_lr(learning_rate)
def __init__(self, prototxt):
super(RMSPropSolver, self).__init__(prototxt=prototxt)
self._optimizer = updaters.RMSPropUpdater(**self._update_param)
self.BuildOptimizer()
def __init__(self, learning_rate, decay, momentum, epsilon=1e-10):
super(RMSPropOptimizer, self).__init__()
self.updater = updaters.RMSPropUpdater(learning_rate, decay, epsilon)
def __init__(self, proto_txt):
super(RMSPropSolver, self).__init__(proto_txt=proto_txt)
self.optimizer = _updaters.RMSPropUpdater(**self._optimizer_arguments)
self.BuildOptimizer()
