def run_task(snapshot_config, *_):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(
snapshot_config=snapshot_config, max_cpus=4, sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = TfEnv(normalize(ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args
)))
# Instantiate the garage objects
policy = GaussianLSTMPolicy(env_spec=env.spec, **policy_args)
baseline = LinearFeatureBaseline(env_spec=env.spec, **baseline_args)
optimizer = ConjugateGradientOptimizer
optimizer_args = {'hvp_approach': FiniteDifferenceHvp(base_eps=1e-5)}
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
optimizer=optimizer,
optimizer_args=optimizer_args,
**algo_args)
sampler_cls = ASTVectorizedSampler
local_runner.setup(
algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args={"open_loop": False,
"sim": sim,
"reward_function": reward_function,
'n_envs': n_parallel})
# Run the experiment
local_runner.train(**runner_args)
def run_task(snapshot_config, *_):
with LocalTFRunner(snapshot_config=snapshot_config, max_cpus=1) as runner:
# Instantiate the example classes
sim = ExampleAVSimulator()
reward_function = ExampleAVReward()
spaces = ExampleAVSpaces()
# Create the environment
env = TfEnv(
normalize(
ASTEnv(blackbox_sim_state=True,
fixed_init_state=True,
s_0=[-0.5, -4.0, 1.0, 11.17, -35.0],
simulator=sim,
reward_function=reward_function,
spaces=spaces)))
# Instantiate the garage objects
policy = GaussianLSTMPolicy(name='lstm_policy',
env_spec=env.spec,
hidden_dim=64)
baseline = LinearFeatureBaseline(env_spec=env.spec)
algo = TRPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
max_path_length=max_path_length,
discount=0.99,
kl_constraint='soft',
max_kl_step=0.01)
sampler_cls = ASTVectorizedSampler
runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args={
"sim": sim,
"reward_function": reward_function
})
runner.train(n_epochs=1, batch_size=4000, plot=False)
print("Installation successfully validated")
def run_task(snapshot_config, *_):
seed = 0
# top_k = 10
np.random.seed(seed)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(snapshot_config=snapshot_config,
max_cpus=4,
sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)
env = TfEnv(env)
policy = ContinuousMLPPolicy(name='ast_agent',
env_spec=env.spec,
**policy_args)
params = policy.get_params()
sess.run(tf.variables_initializer(params))
# Instantiate the garage objects
baseline = ZeroBaseline(env_spec=env.spec)
top_paths = BPQ.BoundedPriorityQueue(**bpq_args)
sampler_cls = ASTVectorizedSampler
sampler_args = {
"open_loop": False,
"sim": sim,
"reward_function": reward_function,
"n_envs": n_parallel
}
if ga_type == 'ga':
print('ga')
algo = GA(env_spec=env.spec,
policy=policy,
baseline=baseline,
top_paths=top_paths,
**algo_args)
elif ga_type == 'gasm':
print('gasm')
algo = GASM(env_spec=env.spec,
policy=policy,
baseline=baseline,
top_paths=top_paths,
**algo_args)
else:
raise NotImplementedError
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
# Run the experiment
local_runner.train(**runner_args)
def run_task(snapshot_config, *_):
seed = 0
# top_k = 10
np.random.seed(seed)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(snapshot_config=snapshot_config,
max_cpus=4,
sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)
top_paths = BPQ.BoundedPriorityQueue(**bpq_args)
if mcts_type == 'mcts':
print('mcts')
algo = MCTS(env=env, top_paths=top_paths, **algo_args)
elif mcts_type == 'mctsbv':
print('mctsbv')
algo = MCTSBV(env=env,
top_paths=top_paths,
**algo_args)
elif mcts_type == 'mctsrs':
print('mctsrs')
algo = MCTSRS(env=env,
top_paths=top_paths,
**algo_args)
else:
raise NotImplementedError
sampler_cls = ASTVectorizedSampler
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args={
"open_loop": False,
"sim": sim,
"reward_function": reward_function,
"n_envs": n_parallel
})
# Run the experiment
local_runner.train(**runner_args)
log_dir = run_experiment_args['log_dir']
with open(log_dir + '/best_actions.p', 'rb') as f:
best_actions = pickle.load(f)
expert_trajectories = []
for actions in best_actions:
sim.reset(s_0=env_args['s_0'])
path = []
for action in actions:
obs = sim.step(action)
state = sim.clone_state()
reward = reward_function.give_reward(
action=action, info=sim.get_reward_info())
path.append({
'state': state,
'reward': reward,
'action': action,
'observation': obs
})
expert_trajectories.append(path)
with open(log_dir + '/expert_trajectory.p', 'wb') as f:
pickle.dump(expert_trajectories, f)
def run_task(snapshot_config, *_):
seed = 0
# top_k = 10
np.random.seed(seed)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(
snapshot_config=snapshot_config, max_cpus=4, sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args
)
top_paths = BPQ.BoundedPriorityQueue(**bpq_args)
if mcts_type == 'mcts':
print('mcts')
algo = MCTS(env=env,
top_paths=top_paths,
**algo_args)
elif mcts_type == 'mctsbv':
print('mctsbv')
algo = MCTSBV(env=env,
top_paths=top_paths,
**algo_args)
elif mcts_type == 'mctsrs':
print('mctsrs')
algo = MCTSRS(env=env,
top_paths=top_paths,
**algo_args)
else:
raise NotImplementedError
sampler_cls = ASTVectorizedSampler
sampler_args['sim'] = sim
sampler_args['reward_function'] = reward_function
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
# Run the experiment
local_runner.train(**runner_args)
log_dir = run_experiment_args['log_dir']
if save_expert_trajectory:
load_convert_and_save_mcts_expert_trajectory(
best_actions_filename=log_dir + '/best_actions.p',
expert_trajectory_filename=log_dir + '/expert_trajectory.p',
sim=sim,
s_0=env_args['s_0'],
reward_function=reward_function)
def run_task(snapshot_config, *_):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
with tf.variable_scope('AST', reuse=tf.AUTO_REUSE):
with LocalTFRunner(snapshot_config=snapshot_config,
max_cpus=4,
sess=sess) as local_runner:
# Instantiate the example classes
sim = ExampleAVSimulator(**sim_args)
reward_function = ExampleAVReward(**reward_args)
spaces = ExampleAVSpaces(**spaces_args)
# Create the environment
if 'id' in env_args:
env_args.pop('id')
env = TfEnv(
normalize(
ASTEnv(simulator=sim,
reward_function=reward_function,
spaces=spaces,
**env_args)))
# Instantiate the garage objects
policy = GaussianLSTMPolicy(env_spec=env.spec,
**policy_args)
baseline = LinearFeatureBaseline(env_spec=env.spec,
**baseline_args)
optimizer = ConjugateGradientOptimizer
optimizer_args = {
'hvp_approach': FiniteDifferenceHvp(base_eps=1e-5)
}
algo = PPO(env_spec=env.spec,
policy=policy,
baseline=baseline,
optimizer=optimizer,
optimizer_args=optimizer_args,
**algo_args)
sampler_cls = ASTVectorizedSampler
sampler_args['sim'] = sim
sampler_args['reward_function'] = reward_function
local_runner.setup(algo=algo,
env=env,
sampler_cls=sampler_cls,
sampler_args=sampler_args)
# Run the experiment
local_runner.train(**runner_args)
if save_expert_trajectory:
load_convert_and_save_drl_expert_trajectory(
last_iter_filename=os.path.join(
run_experiment_args['log_dir'], 'itr_' +
str(runner_args['n_epochs'] - 1) + '.pkl'),
expert_trajectory_filename=os.path.join(
run_experiment_args['log_dir'],
'expert_trajectory.pkl'))
print('done!')