return targets
if __name__=="__main__":
T.set_num_threads(1)
ID = ''.join(random.choices(string.ascii_uppercase + string.digits, k=3))
params = {"iters": 500000, "batchsize": 20, "gamma": 0.99, "lambda_init" : 1., "lambda_min" : 0.3, "policy_lr": 0.0005, "latent_predictor_lr": 0.0005, "weight_decay" : 0.0001, "ppo": True,
"ppo_update_iters": 6, "animate": False, "train" : True,
"note" : "HP, active latent estimation", "ID" : ID}
if socket.gethostname() == "goedel":
params["animate"] = False
params["train"] = True
from src.envs.cartpole_pbt.hangpole import HangPoleBulletEnv
env = HangPoleBulletEnv(animate=params["animate"], latent_input=False, action_input=False)
# Test
if params["train"]:
print("Training")
policy = policies.NN_PG(env, 16, obs_dim=env.obs_dim + env.latent_dim, tanh=False, std_fixed=True)
latent_predictor = policies.RNN_PG(env, hid_dim=8, memory_dim=8, n_temp=2, obs_dim = env.obs_dim + env.act_dim, act_dim=env.latent_dim)
print(params, env.obs_dim, env.act_dim, env.__class__.__name__, policy.__class__.__name__)
train(env, policy, latent_predictor, params)
else:
print("Testing")
policy = T.load('agents/HangPoleBulletEnv_NN_PG_GO5_pg.p')
env.test(policy)
# perlin: P92
# Current experts w/ orientation rew:
# flat: KYH
# holes: 2CW
# tiles: YI7
# triangles: M3X
# Stairs: H1Y
# pipe: W01
# perlin: H03
# Test
if params["train"]:
print("Training")
policy = policies.NN_PG(env, 96)
print(params, env.obs_dim, env.act_dim, env.__class__.__name__, policy.__class__.__name__)
train(env, policy, params)
else:
print("Testing")
policy_name = "MQ7" # LX3: joints + contacts + yaw
policy_path = 'agents/{}_NN_PG_{}_pg.p'.format(env.__class__.__name__, policy_name)
policy = policies.NN_PG(env, 96)
policy.load_state_dict(T.load(policy_path))
env.test(policy, N=10)
print(policy_path)
"weight_decay": 0.001,
"ppo": True,
"ppo_update_iters": 6,
"animate": False,
"train": False,
"env_list": env_list,
"note": "Score, 2E1",
"ID": ID
}
if socket.gethostname() == "goedel":
params["animate"] = False
params["train"] = True
from src.envs.hexapod_trossen_terrain_all import hexapod_trossen_terrain_all as hex_env
env = hex_env.Hexapod(env_list=env_list, max_n_envs=1)
r_lists = []
for i in range(5):
# Test
print("Training: {}/N".format(i + 1))
policy = policies.NN_PG(env, 64, tanh=False, std_fixed=True)
print(params, env.obs_dim, env.act_dim, env.__class__.__name__,
policy.__class__.__name__)
r_list = train(env, policy, params)
r_lists.append(np.array(r_list))
r_lists = np.stack(r_lists, 0)
print(r_lists.shape)
np.save("R_{}_{}".format(env_list, params["ID"]), r_lists)
"animate": True,
"train": False,
"note": "Supervised model learning",
"ID": ID
}
if socket.gethostname() == "goedel":
params["animate"] = False
params["train"] = True
from src.envs.cartpole_pbt.hangpole import HangPoleBulletEnv
env = HangPoleBulletEnv(animate=params["animate"],
latent_input=True,
action_input=False)
# Test
if params["train"]:
print("Training")
model = policies.NN_PG(env,
16,
obs_dim=env.obs_dim,
act_dim=env.obs_dim)
print(params, env.obs_dim, env.act_dim, env.__class__.__name__,
model.__class__.__name__)
train(env, model, params)
else:
print("Testing")
policy_path = 'agents/HangPoleBulletEnv_NN_PG_ETX_pg.p'
policy = T.load(policy_path)
env.test(policy, slow=params["animate"], seed=1338)
print(policy_path)