def main(args=None):
# Parse command line arguments.
parsed_args = parse_command_line_arguments(args)
# Initialize logger.
Logger()
# Initialize dataset object.
init_tf()
dataset_obj = load_dataset(tfrecord_dir=parsed_args.data_dir,
repeat=True,
shuffle_mb=0,
prefetch_mb=100,
max_label_size='full',
verbose=True)
if parsed_args.realism_score: # Compute realism score.
realism_config.datareader = dataset_obj
compute_stylegan_realism(**realism_config)
if parsed_args.truncation_sweep: # Compute truncation sweep.
truncation_config.datareader = dataset_obj
compute_stylegan_truncation(**truncation_config)
peak_gpu_mem_op = tf.contrib.memory_stats.MaxBytesInUse()
peak_gpu_mem_usage = peak_gpu_mem_op.eval()
print('Peak GPU memory usage: %g GB' % (peak_gpu_mem_usage * 1e-9))
def maximum_liklihood():
success = 0
trials = 0
buffer = ExpertBuffer('expert_trajs_point.npz', batch_size=256, number_of_trajs_to_use=20000)
s_mb, a_mb = buffer.sample()
env = PointEnv()
obs = env.reset()
#policy = BCPol(s_mb, a_mb)
def policy_fn(name, ob_space, ac_space, reuse=False):
return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
reuse=reuse, hid_size=32, num_hid_layers=2)
init_tf()
policy = max_liklihood_bc(env,
policy_fn,
buffer,
max_iters=int(1e5),
ckpt_dir=None,
log_dir=None,
task_name=None,
verbose=True)
for test_idx in range(100):
obs = env.reset()
for i in range(250):
act = policy.act(False, obs)[0]
obs, _, _, _ = env.step(act)
# print(obs)
# env.render()
# print(env.get_reward())
rew = env.get_reward()
if np.abs(rew - 1.0) < 0.01:
success += 1
trials += 1
if test_idx % 10 == 0:
print(F"{100*success/trials} percent success rate")
while True:
obs = env.reset()
for i in range(250):
act = policy.act(False, obs)[0] + 0.1*np.random.randn(2)
obs, _, _, _ = env.step(act)
env.render()
def hindsight_supervised_learning():
success = 0
trials = 0
buffer = HindsightBuffer(obs_shape=8, acs_shape=2, future_k=1)
s_mb, a_mb = buffer.sample()
policy = BCPol(s_mb, a_mb)
init_tf()
env = PointEnv(sparse_reward=True)
obs = env.reset()
train_iters = 10000
for i in range(train_iters):
rollout_obs = []
rollout_acts = []
obs = env.reset()
for j in range(250):
act = policy.act(obs)
obs, rew, _, _ = env.step(act)
rollout_obs.append(obs)
rollout_acts.append(act)
buffer.add_traj(rollout_obs, rollout_acts)
rew = env.get_reward()
trials += 1
if np.abs(rew - 1.0) < 0.1:
success += 1
# train policy.
for optim_step in range(10):
s_mb, a_mb = buffer.sample()
loss = policy.train(s_mb, a_mb)
# evaluation
if i % 100 == 0:
#print(F"Loss at step {i} is: {loss}")
print(F"{100*success/trials} percent success rate at step {i}")
success = 0
trials = 0
while True:
for test_idx in range(100):
obs = env.reset()
for i in range(300):
act = policy.act(obs)
obs, _, _, _ = env.step(act)
env.render()
def supervised_learning():
success = 0
trials = 0
buffer = ExpertBuffer('expert_trajs_point.npz')
s_mb, a_mb = buffer.sample()
policy = BCPol(s_mb, a_mb)
init_tf()
env = PointEnv()
obs = env.reset()
train_iters = int(1e5)
for i in range(train_iters):
s_mb, a_mb = buffer.sample()
loss = policy.train(s_mb, a_mb)
if i % 10000 == 0:
#print(F"Loss at step {i} is: {loss}")
#pass
# test the policy.
success = 0
trials = 0
for test_idx in range(10):
obs = env.reset()
for step_idx in range(300):
act = policy.act(obs)
obs, _, _, _ = env.step(act)
# print(obs)
# env.render()
# print(env.get_reward())
rew = env.get_reward()
if np.abs(rew - 1.0) < 0.01:
success += 1
trials += 1
if test_idx == 9:
print(F"{100*success/trials} percent success rate")
while True:
for test_idx in range(100):
obs = env.reset()
for i in range(300):
act = policy.act(obs)
obs, _, _, _ = env.step(act)
env.render()
#! /usr/bin/env python3 from utils import init_tf, output_tf, save_infra_description init_tf() filename = output_tf() save_infra_description(filename)