def debug():
# set up environment and policies
env = TaxiEnvironment()
gamma = 0.98
alpha = 0.6
pi_e = load_taxi_policy("taxi_data/saved_policies/pi19.npy")
pi_other = load_taxi_policy("taxi_data/saved_policies/pi3.npy")
pi_b = MixtureDiscretePolicy(pi_1=pi_e, pi_2=pi_other, pi_1_weight=alpha)
# set up logger
init_state_dist_path = "taxi_data/init_state_dist.npy"
init_state_dist = load_tensor_from_npy(init_state_dist_path).view(-1)
init_state_sampler = DiscreteInitStateSampler(init_state_dist)
logger = DiscreteQLogger(env=env,
pi_e=pi_e,
gamma=gamma,
tensorboard=True,
save_model=True,
init_state_sampler=init_state_sampler)
# generate train and val data
train_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=200000,
burn_in=100000)
val_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=200000,
burn_in=100000)
q = train_q_taxi(env, train_data, val_data, pi_e, pi_b, init_state_sampler,
logger, gamma)
# calculate final performance
policy_val_est = q_estimator(pi_e=pi_e,
gamma=gamma,
q=q,
init_state_sampler=init_state_sampler)
policy_val_oracle = on_policy_estimate(env=env,
pi_e=pi_e,
gamma=gamma,
num_tau=1,
tau_len=1000000)
squared_error = (policy_val_est - policy_val_oracle)**2
print("Test policy val squared error:", squared_error)
def debug():
env = TaxiEnvironment()
gamma = 0.98
alpha = 0.6
pi_e = load_taxi_policy("taxi_data/saved_policies/pi19.npy")
pi_s = load_taxi_policy("taxi_data/saved_policies/pi3.npy")
pi_b = MixtureDiscretePolicy(pi_1=pi_e, pi_2=pi_s, pi_1_weight=alpha)
# set up logger
oracle_tau_len = 1000000
logger = DiscreteWLogger(env=env,
pi_e=pi_e,
pi_b=pi_b,
gamma=gamma,
tensorboard=False,
save_model=False,
oracle_tau_len=oracle_tau_len)
# generate train, val, and test data
tau_len = 200000
burn_in = 100000
train_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=tau_len,
burn_in=burn_in)
val_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=tau_len,
burn_in=burn_in)
# define networks and optimizers
w = StateEmbeddingModel(num_s=env.num_s, num_out=1)
w_lr = 1e-3
w_optimizer = torch.optim.Adam(w.parameters(), lr=w_lr, betas=(0.5, 0.9))
train_w_network_erm_lbfgs(train_data=train_data,
pi_e=pi_e,
pi_b=pi_b,
w=w,
gamma=gamma,
val_data=val_data,
logger=logger)
def debug():
env = TaxiEnvironment()
gamma = 0.98
alpha = 0.6
pi_e = load_taxi_policy("taxi_data/saved_policies/pi19.npy")
pi_other = load_taxi_policy("taxi_data/saved_policies/pi3.npy")
pi_b = MixtureDiscretePolicy(pi_1=pi_e, pi_2=pi_other, pi_1_weight=alpha)
q = StateEmbeddingModel(num_s=env.num_s, num_out=env.num_a)
f = StateEmbeddingModel(num_s=env.num_s, num_out=env.num_a)
# generate train and val data
tau_list = env.generate_roll_out(pi=pi_b, num_tau=1, tau_len=4)
s, a, s_prime, r = tau_list[0]
print("s:", s)
print("a:", a)
print("s_prime:", s_prime)
print("r:", r)
print("")
q_obj, f_obj = q_game_objective(q, f, s, a, s_prime, r, pi_e, gamma)
print("q_obj:", q_obj)
print("f_obj:", f_obj)
def load_eval():
env = TaxiEnvironment(discrete_state=True)
gamma = 0.98
alpha = 0.6
# tau_lens = [50000, 100000, 200000, 400000] # // 10000
tau_len = 200000 # args.tau_len
burn_in = 0 # 100000 # // 10000
init_state_dist_path = "taxi_data/init_state_dist.npy"
pi_e = load_taxi_policy("taxi_data/saved_policies/pi19.npy")
pi_s = load_taxi_policy("taxi_data/saved_policies/pi3.npy")
pi_b = MixtureDiscretePolicy(pi_1=pi_e, pi_2=pi_s, pi_1_weight=alpha)
init_state_dist = load_tensor_from_npy(init_state_dist_path).view(-1)
init_state_sampler = DiscreteInitStateSampler(init_state_dist)
policy_val_oracle = -0.74118131399
print('policy_val_oracle', policy_val_oracle)
print('args.seed', args.seed)
np.random.seed(args.seed)
torch.random.manual_seed(args.seed)
random.seed(args.seed)
train_data = env.generate_roll_out(
pi=pi_b, num_tau=1, tau_len=tau_len, burn_in=burn_in)
train_data_loader = train_data.get_data_loader(1024)
print('train_data generated')
w_path = 'logs/2020-09-13T18:33:35.909858_w_2_200000/290_w_.pt'
q_path = 'logs/2020-09-11T11:04:53.345691_q_2_200000/790_q_.pt'
q_table = torch.zeros((env.num_s, env.num_a))
q = QTableModel(q_table)
q.model.load_state_dict(torch.load(q_path))
w = StateEmbeddingModel(num_s=env.num_s, num_out=1)
w.load_state_dict(torch.load(w_path))
print('[ours] w_oracle and q loaded!')
double_est = double_estimator(
train_data_loader, pi_e, pi_b, w, q, gamma, init_state_sampler)
print('[ours] drl_est', double_est)
def debug():
env = TaxiEnvironment(discrete_state=True)
gamma = 0.98
alpha = 0.6
# tau_lens = [50000, 100000, 200000, 400000] # // 10000
tau_len = args.tau_len
burn_in = 0 # 100000 # // 10000
init_state_dist_path = "taxi_data/init_state_dist.npy"
pi_e = load_taxi_policy("taxi_data/saved_policies/pi19.npy")
pi_s = load_taxi_policy("taxi_data/saved_policies/pi3.npy")
pi_b = MixtureDiscretePolicy(pi_1=pi_e, pi_2=pi_s, pi_1_weight=alpha)
init_state_dist = load_tensor_from_npy(init_state_dist_path).view(-1)
init_state_sampler = DiscreteInitStateSampler(init_state_dist)
policy_val_oracle = -0.74118131399
print('policy_val_oracle', policy_val_oracle)
# for j in tau_lens:
# tau_len = j
# preds = []
# for i in range(100):
# print(j, i)
for i in range(3):
args.seed = i
print('args.seed', args.seed)
np.random.seed(args.seed)
torch.random.manual_seed(args.seed)
random.seed(args.seed)
train_data = env.generate_roll_out(
pi=pi_b, num_tau=1, tau_len=tau_len, burn_in=burn_in)
train_data_loader = train_data.get_data_loader(1024)
val_data = None
print('train_data generated')
now = datetime.datetime.now()
if args.q:
q_path = os.path.join('logs', '_'.join(
[str(now.isoformat()), 'q', str(args.seed), str(args.tau_len)]))
q_logger = SimplestQLogger(
env, pi_e, gamma, init_state_sampler, True, True, q_path, policy_val_oracle)
q = train_q_taxi(env, train_data, val_data, pi_e, pi_b,
init_state_sampler, q_logger, gamma, ERM_epoch=args.q_ERM_epoch, epoch=args.q_epoch)
q_est = q_estimator(pi_e, gamma, q, init_state_sampler)
# SASR = train_data.restore_strcture(discrete=False)
# dataloader doesn't preserve trajectory structure, but SASR does.
# drl1 = double_importance_sampling_estimator(SASR, pi_b, pi_e, args.gamma, q, split_shape=[4, 1, 4, 1])
print('[ours] q_est', q_est)
with open(os.path.join(q_path, 'results.txt'), 'w') as f:
f.write(str(q_est))
print('q_est written in', q_path)
if args.w:
w_path = os.path.join('logs', '_'.join(
[str(now.isoformat()), 'w', str(args.seed), str(args.tau_len)]))
w_logger = SimplestWLogger(
env, pi_e, pi_b, gamma, True, True, w_path, policy_val_oracle)
w = train_w_taxi(env, train_data, val_data, pi_e, pi_b,
init_state_sampler, w_logger, gamma, ERM_epoch=args.w_ERM_epoch, epoch=args.w_epoch)
w_est = w_estimator(train_data_loader, pi_e, pi_b, w)
print('[ours] w_est', w_est)
with open(os.path.join(w_path, 'results.txt'), 'w') as f:
f.write(str(w_est))
print('w_est written in', w_path)
if args.w and args.q:
double_est = double_estimator(
train_data_loader, pi_e, pi_b, w, q, gamma, init_state_sampler)
print('[ours] drl_est', double_est)
with open(os.path.join(q_path, 'results.txt'), 'a') as f:
f.write(str(double_est))
print('double_est written in', q_path)
def debug():
env = TaxiEnvironment(discrete_state=True)
gamma = 0.98
alpha = 0.6
tau_lens = [50000, 100000, 200000, 400000] # // 10000
burn_in = 0 # 100000 # // 10000
init_state_dist_path = "taxi_data/init_state_dist.npy"
pi_e = load_taxi_policy("taxi_data/saved_policies/pi19.npy")
pi_s = load_taxi_policy("taxi_data/saved_policies/pi3.npy")
pi_b = MixtureDiscretePolicy(pi_1=pi_e, pi_2=pi_s, pi_1_weight=alpha)
init_state_dist = load_tensor_from_npy(init_state_dist_path).view(-1)
init_state_sampler = DiscreteInitStateSampler(init_state_dist)
policy_val_oracle = -0.74118131399
print('policy_val_oracle', policy_val_oracle)
for j in tau_lens:
tau_len = j
preds = []
for i in range(100):
print(j, i)
np.random.seed(i)
torch.random.manual_seed(i)
random.seed(i)
train_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=tau_len,
burn_in=burn_in)
train_data_loader = train_data.get_data_loader(1024)
SASR = train_data.restore_strcture(discrete=True)
# print('finished data generation.')
# policy_val_oracle = on_policy_estimate(env=env, pi_e=pi_e, gamma=gamma,
# num_tau=1, tau_len=10000000)
is_est = importance_sampling_estimator(SASR, pi_b, pi_e, gamma)
is_ess = importance_sampling_estimator_stepwise(
SASR, pi_b, pi_e, gamma)
# print('is_est', is_est, 'is_ess', is_ess)
# # Masa's w method, only for discrete
den_discrete = Density_Ratio_discounted(env.n_state, gamma)
_, w_table = train_density_ratio(SASR, pi_b, pi_e, den_discrete,
gamma)
w_table = w_table.reshape(-1)
w = StateEmbeddingModel(num_s=env.num_s, num_out=1)
w.set_weights(torch.from_numpy(w_table))
# print('[Masa] fitted w')
# # Masa's q method is the same as fit_q_tabular
q = fit_q_tabular(train_data, pi_e, gamma)
# print('[Masa] fitted q')
q_est = q_estimator(pi_e, gamma, q, init_state_sampler)
drl_est = double_estimator(train_data_loader, pi_e, pi_b, w, q,
gamma, init_state_sampler)
w_est = w_estimator(train_data_loader, pi_e, pi_b, w)
# print('[Masa] q_est', q_est, '[Masa] w_est',
# w_est, '[Masa] drl_est', drl_est)
preds.append([is_est, is_ess, q_est, drl_est, w_est])
preds = np.array(preds)
errors = (preds - policy_val_oracle)**2
mse = np.mean(errors, axis=0)
print('[is_est, is_ess, q_est, drl_est, w_est] \n', preds)
print('MSE for [is_est, is_ess, q_est, drl_est, w_est] \n', mse)
np.save('estimators/masa_preds_' + str(j), preds)
np.save('estimators/masa_mse_' + str(j), mse)
def debug():
# set up environment and policies
env = TaxiEnvironment(discrete_state=True)
# env = CartpoleEnvironment()
gamma = 0.98
alpha = 0.6
temp = 2.0
hidden_dim = 50
state_dim = 4
pi_e = load_taxi_policy("taxi_data/saved_policies/pi19.npy")
pi_s = load_taxi_policy("taxi_data/saved_policies/pi3.npy")
pi_b = MixtureDiscretePolicy(pi_1=pi_e, pi_2=pi_s, pi_1_weight=alpha)
# pi_e = load_taxi_policy_continuous(
# "taxi_data/saved_policies/pi19.npy", env)
# pi_s = load_taxi_policy_continuous("taxi_data/saved_policies/pi3.npy", env)
# pi_b = GenericMixturePolicy(pi_1=pi_e, pi_2=pi_s, pi_1_weight=alpha)
# set up logger
oracle_tau_len = 100000 # // 10000
init_state_dist_path = "taxi_data/init_state_dist.npy"
init_state_dist = load_tensor_from_npy(init_state_dist_path).view(-1)
init_state_sampler = DiscreteInitStateSampler(init_state_dist)
# init_state_sampler = DecodingDiscreteInitStateSampler(init_state_dist,
# env.decode_state)
# init_state_sampler = CartpoleInitStateSampler(env)
# logger = SimpleDiscretePrintWLogger(env=env, pi_e=pi_e, pi_b=pi_b,
# gamma=gamma,
# oracle_tau_len=oracle_tau_len)
logger = DiscreteWLogger(env=env,
pi_e=pi_e,
pi_b=pi_b,
gamma=gamma,
tensorboard=True,
save_model=True,
oracle_tau_len=oracle_tau_len)
# generate train, val, and test data
tau_len = 200000 # // 10000
burn_in = 100000 # // 10000
train_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=tau_len,
burn_in=burn_in)
print('finish train')
val_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=tau_len,
burn_in=burn_in)
print('finish val')
test_data = env.generate_roll_out(pi=pi_b,
num_tau=1,
tau_len=tau_len,
burn_in=burn_in)
print('finish test')
w = train_w_taxi(env, train_data, val_data, pi_e, pi_b, init_state_sampler,
logger, gamma)
# calculate final performance
test_data_loader = test_data.get_data_loader(1024)
policy_val_est = w_estimator(tau_list_data_loader=test_data_loader,
pi_e=pi_e,
pi_b=pi_b,
w=w)
policy_val_oracle = on_policy_estimate(env=env,
pi_e=pi_e,
gamma=gamma,
num_tau=1,
tau_len=1000000)
squared_error = (policy_val_est - policy_val_oracle)**2
print("Test policy val squared error:", squared_error)