def __init__(self, use_gpu, enable_controller, dim):
self.use_gpu = use_gpu
self.num_of_actions = len(enable_controller)
self.enable_controller = enable_controller
self.dim = dim
print("Initializing Q-Network...")
hidden_dim1 = 64
#hidden_dim1 = 32
hidden_dim2 = 128
hidden_dim3 = 10
hidden_cont = 100
self.model = FunctionSet(
l4=linearL4_link.LinearL4_link(self.dim * self.hist_size *
self.time_M,
hidden_cont,
wscale=np.sqrt(2)),
l5=MU_l6.memory_unit_link(self.dim * self.hist_size * self.time_M,
hidden_dim3 * hidden_cont,
wscale=np.sqrt(2)),
l6=MU_l6.memory_unit_link(self.dim * self.hist_size * self.time_M,
hidden_dim3 * hidden_cont,
wscale=np.sqrt(2)),
l7=attention.Attention(hidden_cont, hidden_dim3 * hidden_cont,
hidden_dim3),
l8=retrieval.Retrieval(hidden_dim3, hidden_dim3 * hidden_cont,
hidden_cont),
l9=F.Bilinear(hidden_cont, hidden_cont, hidden_dim2),
q_value=F.Linear(hidden_dim2,
self.num_of_actions,
initialW=np.zeros(
(self.num_of_actions, hidden_dim2),
dtype=np.float32)))
if self.use_gpu >= 0:
self.model.to_gpu()
self.model_target = copy.deepcopy(self.model)
self.optimizer = optimizers.RMSpropGraves(lr=0.00025,
alpha=0.95,
momentum=0.95,
eps=0.0001)
self.optimizer.setup(self.model.collect_parameters())
# History Data : D=[s(now & 10history), a, r, s_dash, end_episode_flag]
# modified to MQN
self.d = [
np.zeros((self.data_size, self.hist_size * self.time_M, self.dim),
dtype=np.uint8),
np.zeros(self.data_size, dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.int8),
np.zeros((self.data_size, self.hist_size, self.dim),
dtype=np.uint8),
np.zeros((self.data_size, 1), dtype=np.bool)
]
def __init__(self):
super(Linear, self).__init__(l1=F.Bilinear(settings["n_frames"],
settings["n_frames"], 200),
l2=F.Linear(200, 100, wscale=np.sqrt(2)),
l3=F.Linear(100, 100, wscale=np.sqrt(2)),
l4=F.Linear(100, 50, wscale=np.sqrt(2)),
l5=F.Linear(50,
simulator.n_actions,
wscale=np.sqrt(2)))
def setUp(self):
self.f = functions.Bilinear(self.in_shape[0], self.in_shape[1],
self.out_size, True)
self.f.W = numpy.random.uniform(-1, 1,
self.f.W.shape).astype(numpy.float32)
self.f.zero_grads()
self.W = self.f.W.copy()
self.e1 = _uniform(self.batch_size, self.in_shape[0])
self.e2 = _uniform(self.batch_size, self.in_shape[1])
self.gy = _uniform(self.batch_size, self.out_size)
self.y = numpy.einsum('ij,ik,jkl->il', self.e1, self.e2, self.W)
def setUp(self):
self.f = functions.Bilinear(self.in_shape[0], self.in_shape[1],
self.out_size)
self.f.W = _uniform(*self.f.W.shape)
self.f.V1 = _uniform(*self.f.V1.shape)
self.f.V2 = _uniform(*self.f.V2.shape)
self.f.b = _uniform(*self.f.b.shape)
self.f.zero_grads()
self.W = self.f.W.copy()
self.V1 = self.f.V1.copy()
self.V2 = self.f.V2.copy()
self.b = self.f.b.copy()
self.e1 = _uniform(self.batch_size, self.in_shape[0])
self.e2 = _uniform(self.batch_size, self.in_shape[1])
self.gy = _uniform(self.batch_size, self.out_size)
self.y = (numpy.einsum('ij,ik,jkl->il', self.e1, self.e2, self.W) +
self.e1.dot(self.V1) + self.e2.dot(self.V2) + self.b)
def check_invalid(self, initialW, initial_bias, nobias):
with self.assertRaises(AssertionError):
functions.Bilinear(self.in_shape[0], self.in_shape[1],
self.out_size, nobias, initialW, initial_bias)
def check_normal(self, initialW, initial_bias, nobias):
functions.Bilinear(self.in_shape[0], self.in_shape[1], self.out_size,
nobias, initialW, initial_bias)