def test_learning(self):
pol_net = PolNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32)
pol = GaussianPol(self.env.ob_space, self.env.ac_space, pol_net)
targ_pol_net = PolNet(self.env.ob_space, self.env.ac_space, 32, 32)
targ_pol_net.load_state_dict(pol_net.state_dict())
targ_pol = GaussianPol(
self.env.ob_space, self.env.ac_space, targ_pol_net)
qf_net = QNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32)
qf = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space, qf_net)
targ_qf_net = QNet(self.env.ob_space, self.env.ac_space, 32, 32)
targ_qf_net.load_state_dict(targ_qf_net.state_dict())
targ_qf = DeterministicSAVfunc(
self.env.ob_space, self.env.ac_space, targ_qf_net)
sampler = EpiSampler(self.env, pol, num_parallel=1)
optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4)
optim_qf = torch.optim.Adam(qf_net.parameters(), 3e-4)
epis = sampler.sample(pol, max_steps=32)
traj = Traj()
traj.add_epis(epis)
traj = ef.add_next_obs(traj)
traj.register_epis()
result_dict = svg.train(
traj, pol, targ_pol, qf, targ_qf, optim_pol, optim_qf, 1, 32, 0.01, 0.9, 1)
del sampler
def test_learning(self):
ob_space = self.env.real_observation_space
skill_space = self.env.skill_space
ob_skill_space = self.env.observation_space
ac_space = self.env.action_space
ob_dim = ob_skill_space.shape[0] - 4
f_dim = ob_dim
def discrim_f(x): return x
pol_net = PolNet(ob_skill_space, ac_space)
pol = GaussianPol(ob_skill_space, ac_space, pol_net)
qf_net1 = QNet(ob_skill_space, ac_space)
qf1 = DeterministicSAVfunc(ob_skill_space, ac_space, qf_net1)
targ_qf_net1 = QNet(ob_skill_space, ac_space)
targ_qf_net1.load_state_dict(qf_net1.state_dict())
targ_qf1 = DeterministicSAVfunc(ob_skill_space, ac_space, targ_qf_net1)
qf_net2 = QNet(ob_skill_space, ac_space)
qf2 = DeterministicSAVfunc(ob_skill_space, ac_space, qf_net2)
targ_qf_net2 = QNet(ob_skill_space, ac_space)
targ_qf_net2.load_state_dict(qf_net2.state_dict())
targ_qf2 = DeterministicSAVfunc(ob_skill_space, ac_space, targ_qf_net2)
qfs = [qf1, qf2]
targ_qfs = [targ_qf1, targ_qf2]
log_alpha = nn.Parameter(torch.ones(()))
high = np.array([np.finfo(np.float32).max]*f_dim)
f_space = gym.spaces.Box(-high, high, dtype=np.float32)
discrim_net = DiaynDiscrimNet(
f_space, skill_space, h_size=100, discrim_f=discrim_f)
discrim = DeterministicSVfunc(f_space, discrim_net)
optim_pol = torch.optim.Adam(pol_net.parameters(), 1e-4)
optim_qf1 = torch.optim.Adam(qf_net1.parameters(), 3e-4)
optim_qf2 = torch.optim.Adam(qf_net2.parameters(), 3e-4)
optim_qfs = [optim_qf1, optim_qf2]
optim_alpha = torch.optim.Adam([log_alpha], 1e-4)
optim_discrim = torch.optim.SGD(discrim.parameters(),
lr=0.001, momentum=0.9)
off_traj = Traj()
sampler = EpiSampler(self.env, pol, num_parallel=1)
epis = sampler.sample(pol, max_steps=200)
on_traj = Traj()
on_traj.add_epis(epis)
on_traj = ef.add_next_obs(on_traj)
on_traj = ef.compute_diayn_rews(
on_traj, lambda x: diayn_sac.calc_rewards(x, 4, discrim))
on_traj.register_epis()
off_traj.add_traj(on_traj)
step = on_traj.num_step
log_alpha = nn.Parameter(np.log(0.1)*torch.ones(())) # fix alpha
result_dict = diayn_sac.train(
off_traj, pol, qfs, targ_qfs, log_alpha,
optim_pol, optim_qfs, optim_alpha,
step, 128, 5e-3, 0.99, 1, discrim, 4, True)
discrim_losses = diayn.train(
discrim, optim_discrim, on_traj, 32, 100, 4)
del sampler
def test_learning(self):
pol_net = PolNet(self.env.ob_space, self.env.ac_space, h1=32, h2=32)
pol = GaussianPol(self.env.ob_space, self.env.ac_space, pol_net)
qf_net1 = QNet(self.env.ob_space, self.env.ac_space)
qf1 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space,
qf_net1)
targ_qf_net1 = QNet(self.env.ob_space, self.env.ac_space)
targ_qf_net1.load_state_dict(qf_net1.state_dict())
targ_qf1 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space,
targ_qf_net1)
qf_net2 = QNet(self.env.ob_space, self.env.ac_space)
qf2 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space,
qf_net2)
targ_qf_net2 = QNet(self.env.ob_space, self.env.ac_space)
targ_qf_net2.load_state_dict(qf_net2.state_dict())
targ_qf2 = DeterministicSAVfunc(self.env.ob_space, self.env.ac_space,
targ_qf_net2)
qfs = [qf1, qf2]
targ_qfs = [targ_qf1, targ_qf2]
log_alpha = nn.Parameter(torch.zeros(()))
sampler = EpiSampler(self.env, pol, num_parallel=1)
optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4)
optim_qf1 = torch.optim.Adam(qf_net1.parameters(), 3e-4)
optim_qf2 = torch.optim.Adam(qf_net2.parameters(), 3e-4)
optim_qfs = [optim_qf1, optim_qf2]
optim_alpha = torch.optim.Adam([log_alpha], 3e-4)
epis = sampler.sample(pol, max_steps=32)
traj = Traj()
traj.add_epis(epis)
traj = ef.add_next_obs(traj)
traj.register_epis()
result_dict = sac.train(
traj,
pol,
qfs,
targ_qfs,
log_alpha,
optim_pol,
optim_qfs,
optim_alpha,
2,
32,
0.01,
0.99,
2,
)
del sampler
def test_learning(self):
pol_net = PolNet(self.env.observation_space,
self.env.action_space,
h1=32,
h2=32,
deterministic=True)
noise = OUActionNoise(self.env.action_space)
pol = DeterministicActionNoisePol(self.env.observation_space,
self.env.action_space, pol_net,
noise)
targ_pol_net = PolNet(self.env.observation_space,
self.env.action_space,
32,
32,
deterministic=True)
targ_pol_net.load_state_dict(pol_net.state_dict())
targ_noise = OUActionNoise(self.env.action_space)
targ_pol = DeterministicActionNoisePol(self.env.observation_space,
self.env.action_space,
targ_pol_net, targ_noise)
qf_net = QNet(self.env.observation_space,
self.env.action_space,
h1=32,
h2=32)
qf = DeterministicSAVfunc(self.env.observation_space,
self.env.action_space, qf_net)
targ_qf_net = QNet(self.env.observation_space, self.env.action_space,
32, 32)
targ_qf_net.load_state_dict(targ_qf_net.state_dict())
targ_qf = DeterministicSAVfunc(self.env.observation_space,
self.env.action_space, targ_qf_net)
sampler = EpiSampler(self.env, pol, num_parallel=1)
optim_pol = torch.optim.Adam(pol_net.parameters(), 3e-4)
optim_qf = torch.optim.Adam(qf_net.parameters(), 3e-4)
epis = sampler.sample(pol, max_steps=32)
traj = Traj()
traj.add_epis(epis)
traj = ef.add_next_obs(traj)
traj.register_epis()
result_dict = ddpg.train(traj, pol, targ_pol, qf, targ_qf, optim_pol,
optim_qf, 1, 32, 0.01, 0.9)
del sampler
def test_learning(self):
qf_net = QNet(self.env.observation_space, self.env.action_space, 32,
32)
lagged_qf_net = QNet(self.env.observation_space, self.env.action_space,
32, 32)
lagged_qf_net.load_state_dict(qf_net.state_dict())
targ_qf1_net = QNet(self.env.observation_space, self.env.action_space,
32, 32)
targ_qf1_net.load_state_dict(qf_net.state_dict())
targ_qf2_net = QNet(self.env.observation_space, self.env.action_space,
32, 32)
targ_qf2_net.load_state_dict(lagged_qf_net.state_dict())
qf = DeterministicSAVfunc(self.env.observation_space,
self.env.action_space, qf_net)
lagged_qf = DeterministicSAVfunc(self.env.observation_space,
self.env.action_space, lagged_qf_net)
targ_qf1 = CEMDeterministicSAVfunc(self.env.observation_space,
self.env.action_space,
targ_qf1_net,
num_sampling=60,
num_best_sampling=6,
num_iter=2,
multivari=False)
targ_qf2 = DeterministicSAVfunc(self.env.observation_space,
self.env.action_space, targ_qf2_net)
pol = ArgmaxQfPol(self.env.observation_space,
self.env.action_space,
targ_qf1,
eps=0.2)
sampler = EpiSampler(self.env, pol, num_parallel=1)
optim_qf = torch.optim.Adam(qf_net.parameters(), 3e-4)
epis = sampler.sample(pol, max_steps=32)
traj = Traj()
traj.add_epis(epis)
traj = ef.add_next_obs(traj)
traj.register_epis()
result_dict = qtopt.train(traj, qf, lagged_qf, targ_qf1, targ_qf2,
optim_qf, 1000, 32, 0.9999, 0.995, 'mse')
del sampler
pol_net = PolNet(ob_space, ac_space, args.h1, args.h2, deterministic=True)
noise = OUActionNoise(ac_space)
pol = DeterministicActionNoisePol(ob_space, ac_space, pol_net, noise)
targ_pol_net = PolNet(ob_space, ac_space, args.h1, args.h2, deterministic=True)
targ_pol_net.load_state_dict(pol_net.state_dict())
targ_noise = OUActionNoise(ac_space.shape)
targ_pol = DeterministicActionNoisePol(
ob_space, ac_space, targ_pol_net, targ_noise)
qf_net = QNet(ob_space, ac_space, args.h1, args.h2)
qf = DeterministicSAVfunc(ob_space, ac_space, qf_net)
targ_qf_net = QNet(ob_space, ac_space, args.h1, args.h2)
targ_qf_net.load_state_dict(qf_net.state_dict())
targ_qf = DeterministicSAVfunc(ob_space, ac_space, targ_qf_net)
sampler = EpiSampler(env, pol, num_parallel=args.num_parallel, seed=args.seed)
optim_pol = torch.optim.Adam(pol_net.parameters(), args.pol_lr)
optim_qf = torch.optim.Adam(qf_net.parameters(), args.qf_lr)
off_traj = Traj(args.max_steps_off)
total_epi = 0
total_step = 0
max_rew = -1e6
while args.max_epis > total_epi:
with measure('sample'):
set_device(device)
score_file = os.path.join(args.log, 'progress.csv')
logger.add_tabular_output(score_file)
env = GymEnv(args.env_name,
log_dir=os.path.join(args.log, 'movie'),
record_video=args.record)
env.env.seed(args.seed)
observation_space = env.observation_space
action_space = env.action_space
qf_net = QNet(observation_space, action_space, args.h1, args.h2)
lagged_qf_net = QNet(observation_space, action_space, args.h1, args.h2)
lagged_qf_net.load_state_dict(qf_net.state_dict())
targ_qf1_net = QNet(observation_space, action_space, args.h1, args.h2)
targ_qf1_net.load_state_dict(qf_net.state_dict())
targ_qf2_net = QNet(observation_space, action_space, args.h1, args.h2)
targ_qf2_net.load_state_dict(lagged_qf_net.state_dict())
qf = DeterministicSAVfunc(observation_space,
action_space,
qf_net,
data_parallel=args.data_parallel)
lagged_qf = DeterministicSAVfunc(observation_space,
action_space,
lagged_qf_net,
data_parallel=args.data_parallel)
targ_qf1 = CEMDeterministicSAVfunc(observation_space,
action_space,
targ_qf1_net,
pol_net = PolNet(ob_space, ac_space)
pol = GaussianPol(ob_space,
ac_space,
pol_net,
data_parallel=args.data_parallel,
parallel_dim=0)
qf_net1 = QNet(ob_space, ac_space)
qf1 = DeterministicSAVfunc(ob_space,
ac_space,
qf_net1,
data_parallel=args.data_parallel,
parallel_dim=0)
targ_qf_net1 = QNet(ob_space, ac_space)
targ_qf_net1.load_state_dict(qf_net1.state_dict())
targ_qf1 = DeterministicSAVfunc(ob_space,
ac_space,
targ_qf_net1,
data_parallel=args.data_parallel,
parallel_dim=0)
qf_net2 = QNet(ob_space, ac_space)
qf2 = DeterministicSAVfunc(ob_space,
ac_space,
qf_net2,
data_parallel=args.data_parallel,
parallel_dim=0)
targ_qf_net2 = QNet(ob_space, ac_space)
targ_qf_net2.load_state_dict(qf_net2.state_dict())
targ_qf2 = DeterministicSAVfunc(ob_space,
# Gymのenviromentを生成
from pybullet_envs.bullet.racecarGymEnv import RacecarGymEnv
env = RacecarGymEnv(renders=False, isDiscrete=False)
# 観測と行動の次元
observation_space = env.observation_space
action_space = env.action_space
# Q-Network
qf_net = QNet(observation_space, action_space, args.h1, args.h2)
qf = DeterministicSAVfunc(
observation_space, action_space, qf_net,
data_parallel=args.data_parallel) # 決定的行動状態価値関数?q-netの出力の形を少し整える
# target Q network theta1
targ_qf1_net = QNet(observation_space, action_space, args.h1, args.h2)
targ_qf1_net.load_state_dict(qf_net.state_dict()) # model（重み）をロード(q-netからコピー)
targ_qf1 = CEMDeterministicSAVfunc(
observation_space,
action_space,
targ_qf1_net,
num_sampling=args.num_sampling,
num_best_sampling=args.num_best_sampling,
num_iter=args.num_iter,
multivari=args.multivari,
data_parallel=args.data_parallel,
save_memory=args.save_memory) #CrossEntropy Methodよくわからん
# lagged network
lagged_qf_net = QNet(observation_space, action_space, args.h1, args.h2)
lagged_qf_net.load_state_dict(
qf_net.state_dict()) # model（重み）をロード(theta1からコピー)
def main():
pygame.init() # 初期化
(w, h) = (480, 320)
screen = pygame.display.set_mode((w, h), FULLSCREEN) # window size
pygame.display.set_caption("Sikamaru") # window bar
# initialization
tx = 0
ty = 0
sika = Sikamaru((w / 2, h / 2))
sleep_count = 5
eat_mode = 100
esa = Food()
wait = True
seed = 42
# TODO define RL agent
'''
state : 4D (sikaposi, esaposi)
action : 2D (-20,+20)^2
SAC
simple_net : 30,30
'''
np.random.seed(seed)
torch.manual_seed(seed)
low = np.zeros(4)
high = w * np.ones(4)
ob_space = gym.spaces.Box(low=low, high=high)
ac_space = gym.spaces.Discrete(4)
ac_dict = {
0: np.array([-20, 0]),
1: np.array([20, 0]),
2: np.array([0, -20]),
3: np.array([0, 20])
}
pol_net = PolNet(ob_space, ac_space)
pol = CategoricalPol(ob_space, ac_space, pol_net)
qf_net1 = QNet(ob_space, ac_space)
qf1 = DeterministicSAVfunc(ob_space, ac_space, qf_net1)
targ_qf_net1 = QNet(ob_space, ac_space)
targ_qf_net1.load_state_dict(qf_net1.state_dict())
targ_qf1 = DeterministicSAVfunc(ob_space, ac_space, targ_qf_net1)
qf_net2 = QNet(ob_space, ac_space)
qf2 = DeterministicSAVfunc(ob_space, ac_space, qf_net2)
targ_qf_net2 = QNet(ob_space, ac_space)
targ_qf_net2.load_state_dict(qf_net2.state_dict())
targ_qf2 = DeterministicSAVfunc(ob_space, ac_space, targ_qf_net2)
qfs = [qf1, qf2]
targ_qfs = [targ_qf1, targ_qf2]
log_alpha = nn.Parameter(torch.ones(()))
optim_pol = torch.optim.Adam(pol_net.parameters(), 1e-4)
optim_qf1 = torch.optim.Adam(qf_net1.parameters(), 3e-4)
optim_qf2 = torch.optim.Adam(qf_net2.parameters(), 3e-4)
optim_qfs = [optim_qf1, optim_qf2]
optim_alpha = torch.optim.Adam([log_alpha], 1e-4)
# off_traj = Traj()
while (True):
screen.fill((
0,
100,
0,
)) # backgroud color
# my procedure
## env
obs = make_obs((tx, ty), sika.posi, w, h)
ac_real, ac, a_i = pol.deterministic_ac_real(
torch.tensor(obs, dtype=torch.float))
# ac_real = ac_real.reshape(pol.ac_space.shape)
a = rule_act((tx, ty), sika.posi)
# a = ac_dict[int(ac_real)]
nx = sika.posi[0] + a[0]
nx = max(min(nx, w), 0)
ny = sika.posi[1] + a[1]
ny = max(min(ny, h), 0)
sika.move((nx, ny))
screen.blit(sika.get_im(), sika.rect)
if esa.life: # RL
# TOOD:record as epi
screen.blit(esa.im, esa.rect)
# scr
rew = esa.life_step(sika)
if rew > 0:
sika.bigup()
if esa.life == 0:
pass
#TODO add one epi and learn
wait = False
if wait:
pygame.time.wait(500)
wait = True
pygame.display.update() # 画面更新
## event
for event in pygame.event.get():
if event.type == MOUSEBUTTONDOWN and event.button == 1:
tx, ty = event.pos
esa.set((tx, ty))
if event.type == KEYDOWN:
if event.key == K_ESCAPE:
sys.exit()
if event.type == QUIT: # 終了処理
pygame.quit()
sys.exit()