def run(num_envs=16,
hidden_dim=256,
batch_size=1024,
iterations=1000,
log_interval=10,
runs=1):
envs = [tl.make_nh_waypoint_3d() for i in range(num_envs)]
envs = SubprocVecEnv(envs)
t_env = tenv.WaypointEnv3D()
state_dim = t_env.observation_space.shape[0]
action_dim = t_env.action_space.shape[0]
path = os.getcwd() + "/nh_waypoint_3d/"
for i in range(runs):
agent = ag.Agent(state_dim, hidden_dim, action_dim, dim=3)
opt = torch.optim.Adam(agent.parameters(), lr=1e-4)
ep, rew, agent = tl.train_mp(envs,
t_env,
agent,
opt,
batch_size,
iterations,
log_interval,
render=False,
fname=path + "gaussian_" + str(2))
if i == 0:
csv_input = pd.DataFrame()
csv_input["timesteps"] = ep
csv_input["run" + str(i)] = rew
csv_input.to_csv(path + "data.csv", index=False)
def run(hidden_dim=256):
path = os.getcwd()+"/nh_waypoint_3d/"
fname = path+"gaussian_2_term.pth.tar"
t_env = tenv.WaypointEnv3D()
state_dim = t_env.observation_space.shape[0]
action_dim = t_env.action_space.shape[0]
agent = ag.Agent(state_dim, hidden_dim, action_dim, dim=3)
print("Agent initialized, loading state dictionary.")
agent.load_state_dict(torch.load(fname, map_location=lambda storage, loc: storage))
print()
print("State dictionary loaded")
k = [tl.test(t_env, agent, render=True) for _ in range(100)]
rewards = sum(k)/len(k)
print("Mean reward: ", rewards)
def run(hidden_dim=256):
path = os.getcwd() + "/_2d/soft_2d/"
fname = path + "3-wps_.pth.tar"
t_env = tenv.TrajectoryEnv2D()
state_dim = t_env.observation_space.shape[0]
action_dim = t_env.action_space.shape[0]
agent = ag.Agent(state_dim, hidden_dim, action_dim, dim=2, lookahead=3)
print("Agent initialized, loading state dictionary.")
agent.load_state_dict(
torch.load(fname, map_location=lambda storage, loc: storage))
print()
print("State dictionary loaded")
k = [tl.test(t_env, agent, render=True) for _ in range(100)]
rewards = sum(k) / len(k)
print("Mean reward: ", rewards)
def run(num_envs=16, hidden_dim=256, batch_size=1024, iterations=1000, log_interval=10, runs=3):
envs = [tl.make_traj_2d() for i in range(num_envs)]
envs = SubprocVecEnv(envs)
t_env = tenv.TrajectoryEnv2D()
t_env.num_fut_wp = int(cfg.waypoints-1)
state_size = 10+7*(t_env.num_fut_wp+1)
t_env.observation_space = gym.spaces.Box(-1, 1, shape=(state_size,))
state_dim = t_env.observation_space.shape[0]
action_dim = t_env.action_space.shape[0]
path = os.getcwd()+"/_2d/traj_2d/"
for i in range(runs):
agent = ag.Agent(state_dim, hidden_dim, action_dim, dim=2, lookahead=lookahead)
opt = torch.optim.Adam(agent.parameters(), lr=cfg.lr)
ep, rew, agent = tl.train_mp(envs, t_env, agent, opt, batch_size, iterations, log_interval, render=False, fname=path+wps+"-wps_")
if i == 0:
csv_input = pd.DataFrame()
csv_input["timesteps"] = ep
csv_input["run"+str(i)] = rew
csv_input.to_csv(path+"data_wp-"+wps+".csv", index=False)
def run(hidden_dim=256):
path = os.getcwd() + "/waypoint_3d/"
fname = path + "gaussian_2_term.pth.tar"
t_env = tenv.WaypointEnv3D()
state_dim = t_env.observation_space.shape[0]
action_dim = t_env.action_space.shape[0]
agent = ag.Agent(state_dim, hidden_dim, action_dim, dim=3)
print("Agent initialized, loading state dictionary.")
agent.load_state_dict(
torch.load(fname, map_location=lambda storage, loc: storage))
print()
print("State dictionary loaded")
uvws, pqrs, ts = test(t_env, agent)
fig = plt.figure(figsize=(5, 5))
ax = fig.add_subplot(111)
ax.plot(ts, uvws)
ax.set_xlabel('time (s)')
ax.set_ylabel('u (m/s)')
ax.set_xlim([0, 3])
ax.set_ylim([-1.5, 1.5])
plt.savefig('./figures/value_density.png')
print("figure saved")
def run(hidden_dim=256):
path = os.getcwd() + "/waypoint_2d/"
fname = path + "gaussian_2_term.pth.tar"
t_env = tenv.WaypointEnv2D()
state_dim = t_env.observation_space.shape[0]
action_dim = t_env.action_space.shape[0]
agent = ag.Agent(state_dim, hidden_dim, action_dim)
print("Agent initialized, loading state dictionary.")
agent.load_state_dict(
torch.load(fname, map_location=lambda storage, loc: storage))
print()
print("State dictionary loaded")
xs = np.linspace(0, 3, 30)
ys = np.linspace(-1.5, 1.5, 30)
XS, YS = np.meshgrid(xs, ys)
VALUE = np.zeros(XS.shape)
for i, x in enumerate(xs):
for j, y in enumerate(ys):
arr = np.array([x, y])
state = torch.Tensor(arr).to(device)
value = agent.get_integrated_value(state).item()
VALUE[i, j] = exp(-value)
fig = plt.figure(figsize=(5, 5))
ax = fig.add_subplot(111)
ax.pcolormesh(XS, YS, VALUE, cmap="plasma")
ax.set_xlabel('body x distance (m)')
ax.set_ylabel('body y distance (m)')
ax.set_xlim([0, 3])
ax.set_ylim([-1.5, 1.5])
plt.savefig('./figures/value_density.png')
print("figure saved")