def test_call_hooks_uninitialized_param(self):
target = UninitializedChain()
opt = optimizers.MomentumSGD()
opt.setup(target)
opt.add_hook(optimizer.Lasso(rate=0.0005))
target(np.ones((4, 10), dtype=np.float32))
opt.call_hooks()
def check_lasso(self):
w = self.target.param.data
g = self.target.param.grad
xp = cuda.get_array_module(w)
decay = 0.2
expect = w - g - decay * xp.sign(w)
opt = optimizers.SGD(lr=1)
opt.setup(self.target)
opt.add_hook(optimizer.Lasso(decay))
opt.update()
testing.assert_allclose(expect, w)
def __init__(self,
agent,
memory_size=10**4,
replay_size=32,
gamma=0.99,
initial_exploration=10**4,
target_update_freq=10**4,
learning_rate=0.00025,
epsilon_decay=1e-6,
minimum_epsilon=0.1,
L1_rate=None):
self.agent = agent
self.target = Q(self.agent.q.n_history,
self.agent.q.n_action,
on_gpu=self.agent.q.on_gpu)
self.memory_size = memory_size
self.replay_size = replay_size
self.gamma = gamma
self.initial_exploration = initial_exploration
self.target_update_freq = target_update_freq
self.laerning_rate = learning_rate
self.epslon_decay = epsilon_decay
self.minimum_epsilon = minimum_epsilon
self._step = 0
# prepare for replay
n_hist = self.agent.q.n_history
size = self.agent.q.SIZE
self.memory = [
np.zeros((memory_size, n_hist, 3, size, size), dtype=np.float32),
np.zeros(memory_size, dtype=np.uint8),
np.zeros((memory_size, 1), dtype=np.float32),
np.zeros((memory_size, n_hist, 3, size, size), dtype=np.float32),
np.zeros((memory_size, 1), dtype=np.bool)
]
self.memory_text = [
"state", "action", "reward", "next_state", "episode_end"
]
#prepare optimizer
self.optimizer = optimizers.RMSpropGraves(lr=learning_rate,
alpha=0.95,
momentum=0.95,
eps=0.01)
self.optimizer.setup(self.agent.q)
if L1_rate is not None:
self.optimizer.add_hook(optimizer.Lasso(L1_rate))
self._loss = 9
self._qv = 0
trs = False
train_iter = iterators.SerialIterator(dataset_train,
batch_size=args.batchsize,
shuffle=trs)
if args.numval > 0:
val_iter = iterators.SerialIterator(dataset_val,
batch_size=len(dataset_val),
repeat=False,
shuffle=False)
# -- Set optimizers
optimizer1 = use_optimizer(lr=args.learning_rate)
optimizer1.setup(loss.phi)
optimizer2 = use_optimizer(lr=args.learning_rate)
optimizer2.setup(loss.net)
optimizer1.add_hook(optimizer_module.Lasso(args.beta))
optimizer2.add_hook(optimizer_module.WeightDecay(args.gamma))
# -- Set a trigger
if args.log_in_iteration:
trigger = (1, 'iteration')
else:
trigger = (1, 'epoch')
# -- Set a trainer
if args.fixed_embedder:
optimizer_dict = {'net': optimizer2}
else:
optimizer_dict = {'phi': optimizer1, 'net': optimizer2}
updater = lkis.Updater(train_iter,
optimizer_dict,