def learn_on_ppo1(env_name,
num_timesteps,
seed,
model_path,
lognum,
old_model_path=None):
log_dir = str(Path.cwd().parent.parent.joinpath('logs', str(lognum)))
# win: tensorboard --logdir="E:\gitstuff\bachelor_thesis_snake\logs\x"
logger.configure(
dir=log_dir,
format_strs=['stdout', 'log', 'csv', 'json', 'tensorboard'])
# mac: go to cwd, pipenv shell, tensorboard --logdir=/Users/julian.schmitz/gitstuff/bachelor_thesis_snake/logs/1
sess = U.make_session(num_cpu=1)
sess.__enter__()
set_global_seeds(seed)
env = gym.make(env_name)
env = bench.Monitor(env, os.path.join(logger.get_dir(), "monitor.json"))
pi = None
if old_model_path is not None:
pi = cnn_policy.CnnPolicy('pi', env.observation_space,
env.action_space)
# tf.train.Saver().restore(sess, old_model_path) # TODO does that load the model and we can go on?
tf.train.Saver().restore(
tf.get_default_session(), old_model_path
) # TODO does that load the model and we can learn on?
my_pposgd_simple.learn(
env,
policy_fn=policy_fn_cnn,
max_timesteps=num_timesteps,
# max_episodes=1000,
timesteps_per_actorbatch=2048, # TODO or 4096 when obstacles
clip_param=0.2,
entcoeff=0.0,
optim_epochs=10,
optim_stepsize=3e-4,
optim_batchsize=64,
gamma=0.99,
lam=0.95,
schedule='linear',
callback=callback,
pi=pi)
# env.close()
saver = tf.train.Saver()
saver.save(sess, model_path)
print('saved to ' + model_path)
def policy_fn(name, ob_space, ac_space): #pylint: disable=W0613
return cnn_policy.CnnPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space)
def policy_fn_cnn(name, ob_space, ac_space):
return cnn_policy.CnnPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
kind='small') # TODO kind='large'
def policy_fn(name, ob_space, ac_space):
return cnn_policy.CnnPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space)
def policy_fn(name, ob_space, ac_space, sess=None, placeholders=None): # pylint: disable=W0613
return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space, sess=sess,
placeholders=placeholders)
def policy_fn(name, ob_space, ac_space):
"""Given an obs, returns an act."""
return cnn_policy.CnnPolicy(name=name, ob_space=ob_space,
ac_space=ac_space)
