l1, b1, l2, b2 = model.get_gen_weights()
weight_size = np.sum(
list(l1.flatten().shape) + list(b1.flatten().shape) +
list(l2.flatten().shape) + list(b2.flatten().shape))
dis = models.DiscriminatorZ(args, obs_size, n_actions)
dis = dis.to(device)
# ms-CEM policy (refers to the guiding CEM policy)
dim_theta = (obs_size + 1) * n_actions
# Initialize mean and standard deviation
theta_mean = np.zeros(dim_theta)
theta_std = np.ones(dim_theta)
theta = np.random.multivariate_normal(theta_mean, np.diag(theta_std**2))
if LAYER == 2: # 2-layer network
msCEM_model = models.NetW_2Layer(args, obs_size, n_actions)
if LAYER == 3: # 3-layer network
msCEM_model = models.NetW_3Layer(args, obs_size, n_actions)
msCEM_model = msCEM_model.to(device)
criterionGen = nn.BCELoss()
criterionDis = nn.BCELoss()
criterionMSE = nn.MSELoss()
print(model)
print(dis)
print("Observation Size: {} \t Action Size: {}".format(
obs_size, n_actions))
print("Model_param: {}".format(utils.count_parameters(model)))
print("Encoder_param: {}".format(utils.count_parameters(model.encoder)))
device = torch.device("cuda:" + DEVICE if torch.cuda.is_available() else "cpu")
if not use_cuda:
device = torch.device("cpu")
print(device)
dim_theta = (obs_size + 1) * n_actions
# Initialize mean and standard deviation
theta_mean = np.zeros(dim_theta)
theta_std = np.ones(dim_theta)
# main policy network
theta = np.random.multivariate_normal(theta_mean, np.diag(theta_std ** 2))
if LAYER == 2:
model = models.NetW_2Layer(args, obs_size, n_actions) # 2-layer
if LAYER == 3:
model = models.NetW_3Layer(args, obs_size, n_actions) # 3-layer
model = model.to(device)
print("Observation Size: {} \t Action Size: {}".format(obs_size, n_actions))
start_time = time.time()
# define agent start state
agent_start_list = [(INI_JOING_ANGLES, [369,430])]
e_batch_buffer = utils.elite_batch_bufferW_msCEM(25) # 10
global_eList = []
total_test_reward_accum = 0