def setup_experiment(
sess,
agent_config,
env_config,
paths,
seed=None
):
"""
Initialize an experiment
sess (tf.Session)
agent_config (dict)
env_config (dict)
paths (dict)
seed (int)
"""
env = energypy.make_env(**env_config)
save_args(env_config, path=paths['env_args'])
if seed:
logger.info('random seed is {}'.format(seed))
env.seed(seed)
agent_config['env'] = env
agent_config['sess'] = sess
agent_config['act_path'] = paths['tb_act']
agent_config['learn_path'] = paths['tb_learn']
agent_memory = agent_config.pop('load_memory', None)
if agent_memory:
agent_config['load_memory_path'] = paths['memory']
agent = energypy.make_agent(**agent_config)
save_args(agent_config, path=paths['agent_args'])
if hasattr(agent, 'acting_writer'):
agent.acting_writer.add_graph(sess.graph)
# TODO copy the dataset into the run folder as well
logger.info('setup experiment of {} steps'.format(total_steps))
return agent, env
def setup_agent(sess, double_q=False):
"""
Sets up an agent & fills memory
args
sess (tf.Session)
returns
agent (energypy DQN agent)
env (energypy Battery environment)
"""
env = energypy.make_env('2048', observation_dims='2D')
# use high learning rate to get weight changes
agent = energypy.make_agent(agent_id='dqn',
sess=sess,
env=env,
total_steps=10,
discount=0.9,
memory_type='deque',
learning_rate=1.0,
double_q=double_q,
update_target_net=100,
network='conv',
filters=(8, 16),
kernels=(2, 2),
strides=(1, 1))
for step in range(48):
obs = env.observation_space.sample()
action = env.action_space.sample()
reward = random.random() * 10
next_obs = env.observation_space.sample()
done = random.choice([True, False])
agent.remember(obs, action, reward, next_obs, done)
batch = agent.memory.get_batch(agent.batch_size)
return agent, batch, env
def setup_run(cfg, run, sess):
run_cfg = make_run_config(cfg, run)
run_logger = make_new_logger('run_setup', run_cfg['run_dir'])
runner = Runner(sess, run_cfg)
dump_config(run_cfg, run_logger)
env_config = run_cfg['env']
env = make_env(**env_config)
if hasattr(env.observation_space, 'info') and hasattr(
env.state_space, 'info'):
run_logger.debug(json.dumps({'state_info': env.state_space.info}))
run_logger.debug(
json.dumps({'observation_info': env.observation_space.info}))
agent_config = run_cfg['agent']
agent_config['env'] = env
agent_config['sess'] = sess
agent_config['tensorboard_dir'] = run_cfg['tensorboard_dir']
agent = make_agent(**agent_config)
return run_cfg, agent, env, runner