EPS = 1e-8 # numerical residual
TEST_EP = 10
# ppo-penalty
KL_TARGET = 0.01
LAM = 0.5
# ppo-clip
EPSILON = 0.2
RENDER = False
PLOT_RESULT = False
ARG_NAME = 'PPO'
METHOD = ['penalty', 'clip'][0]
# underlying controller candidates
model_1 = Actor(state_size=4, action_size=1, seed=0).to(device)
model_1.load_state_dict(torch.load("./actor5000_1.pth"))
model_1.eval()
model_2 = Actor(state_size=4, action_size=1, seed=0).to(device)
model_2.load_state_dict(torch.load("./actor4850_1.pth"))
model_2.eval()
WEIGHT = 1
# assert False
class AddBias(nn.Module):
def __init__(self, bias):
super(AddBias, self).__init__()
self._bias = nn.Parameter(bias.unsqueeze(1))
Variable = lambda *args, **kwargs: autograd.Variable(*args, **kwargs).cuda(
) if USE_CUDA else autograd.Variable(*args, **kwargs)
batch_size = 128
gamma = 0.99
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 3000
replay_buffer = ReplayBuffer(int(5e3))
epsilon_by_frame = lambda frame_idx: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model_1 = Actor(state_size=2,
action_size=1,
seed=0,
fc1_units=25,
fc2_units=None).to(device)
model_1.load_state_dict(torch.load("./models/actor_2800.pth"))
model_1.eval()
model_2 = Actor(state_size=2, action_size=1, seed=0, fc1_units=25).to(device)
if EXP1:
model_2.load_state_dict(torch.load("./0731actors/actor_2400.pth"))
else:
model_2.load_state_dict(torch.load("./0801actors/actor_1400.pth"))
model_2.eval()
Individual = Individualtanh(state_size=2, action_size=1, seed=0).to(device)
agent = Agent(state_size=2,
num_warmup = 1000
num_train = 200000
num_eval = 0
buffer_length = 600000
# env = NormalizedEnv(gym.make('Pendulum-v0'))
GODOT_BIN_PATH = "InvPendulum/InvPendulum.x86_64"
env_abs_path = "InvPendulum/InvPendulum.pck"
env = NormalizedEnv(
InvPendulumEnv(exec_path=GODOT_BIN_PATH,
env_path=env_abs_path,
render=True))
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
policy = Actor(num_states, num_actions)
policy.load_state_dict(torch.load('./policy.pkl'))
state = env.reset()
state = state.to(dtype=torch.float32)
traced_policy = torch.jit.trace(policy, state)
print(traced_policy.graph)
print(traced_policy.code)
traced_policy.save('ddpg_policy.jit')
for step in range(1000):
action = policy(state)
# torch.tensor([1.0 for i in range(num_actions)])).sample().to(device='cuda')
time.sleep(0.02)