def test_init(self):
ddpg, locals = self.create_ddpg()
env_spec = locals['env_spec']
env = locals['env']
mlp_dyna = self.create_continuous_mlp_global_dynamics_model(env_spec=env_spec)[0]
algo = self.create_dyna(env_spec=env_spec, model_free_algo=ddpg, dyanmics_model=mlp_dyna)[0]
algo.init()
st = env.reset()
data = TransitionData(env_spec)
for _ in range(100):
ac = algo.predict(st)
new_st, re, done, _ = env.step(action=ac)
data.append(state=st,
new_state=new_st,
reward=re,
action=ac,
done=done)
algo.append_to_memory(samples=data)
pre_res = 10000
for i in range(20):
print(algo.train(batch_data=data))
print(algo.train(batch_data=data, state='state_dynamics_training'))
print(algo.train(batch_data=data, state='state_agent_training'))
res = algo.test_dynamics(env=env, sample_count=100)
self.assertLess(list(res.values())[0], pre_res)
self.assertLess(list(res.values())[1], pre_res)
print(res)
algo.test()
def test_with_dqn(self):
dqn, local = self.create_dqn()
env = local['env']
env_spec = local['env_spec']
dqn.init()
st = env.reset()
from baconian.common.sampler.sample_data import TransitionData
a = TransitionData(env_spec)
res = []
for i in range(100):
ac = dqn.predict(obs=st, sess=self.sess, batch_flag=False)
st_new, re, done, _ = env.step(action=ac)
a.append(state=st,
new_state=st_new,
action=ac,
done=done,
reward=re)
dqn.append_to_memory(a)
res.append(
dqn.train(batch_data=a,
train_iter=10,
sess=None,
update_target=True)['average_loss'])
res.append(
dqn.train(batch_data=None,
train_iter=10,
sess=None,
update_target=True)['average_loss'])
print(dqn._status())
print(dqn._status._info_dict_with_sub_info)
def test_trajectory_data(self):
env = make('Acrobot-v1')
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
a = TrajectoryData(env_spec)
tmp_traj = TransitionData(env_spec)
st = env.reset()
re_list = []
st_list = []
for i in range(100):
ac = env_spec.action_space.sample()
st_new, re, done, _ = env.step(action=ac)
st_list.append(st_new)
re_list.append(re)
if (i + 1) % 10 == 0:
done = True
else:
done = False
tmp_traj.append(state=st,
new_state=st_new,
action=ac,
done=done,
reward=re)
if done:
a.append(tmp_traj.get_copy())
tmp_traj.reset()
self.assertEqual(a.trajectories.__len__(), 10)
for traj in a.trajectories:
self.assertEqual(len(traj), 10)
def DynaMLP_get_action(self, mlp_dyna: DynamicsModel, env: Env, state,
cost_fn, num_simulated_paths, horizon):
'''
mpc.ModelBasedModelPredictiveControl.predict()
:param mlp_dyna:
:param env:
:param state:
:param cost_fn:
:param num_simulated_paths:
:param horizon:
:return:
'''
rollout = TrajectoryData(env_spec=env.env_spec)
for i in range(num_simulated_paths):
path = TransitionData(env_spec=env.env_spec)
obs = state
for j in range(horizon):
action = env.action_space.sample()
obs_ = mlp_dyna.step(action=action, state=obs)
cost = cost_fn(obs, action, obs_)
path.append(obs, action, obs_, False, -cost)
obs = obs_
rollout.append(path)
rollout.trajectories.sort(key=lambda x: x.cumulative_reward,
reverse=True)
optimial_action = rollout.trajectories[0].action_set[0]
return optimial_action
def test_sample_batch(self):
env = make('ModifiedHalfCheetah')
env.init()
env_spec = env.env_spec
random_buffer = TransitionData(env_spec=env_spec, obs_shape=env_spec.obs_shape, \
action_shape=env_spec.action_shape, size=100)
print("====> Random Sample")
num_trajectory = 1
max_step = 30
for i in range(num_trajectory):
ep_len = 0
obs = env.reset()
while ep_len < max_step:
act = self.RandomController_get_action(env=env, state=obs)
obs_, reward, done, _ = env.step(act)
random_buffer.append(obs, act, obs_, done, reward)
assert not done
obs = obs_
ep_len += 1
batch_data_1 = random_buffer.sample_batch(batch_size=16,
shuffle_flag=True)
assert isinstance(batch_data_1, dict)
print(batch_data_1.keys())
self.assertEqual(len(batch_data_1['action_set']), 16)
def sample(self, batch_size) -> SampleData:
if self.nb_entries < batch_size:
raise MemoryBufferLessThanBatchSizeError()
# todo This will be changed to prioritised
batch_idxs = np.random.randint(self.nb_entries - 2, size=batch_size)
obs0_batch = self.observations0.get_batch(batch_idxs)
obs1_batch = self.observations1.get_batch(batch_idxs)
action_batch = self.actions.get_batch(batch_idxs)
reward_batch = self.rewards.get_batch(batch_idxs)
terminal1_batch = self.terminals1.get_batch(batch_idxs)
result = {
'obs0': array_min2d(obs0_batch),
'obs1': array_min2d(obs1_batch),
'rewards': array_min2d(reward_batch),
'actions': array_min2d(action_batch),
'terminals1': array_min2d(terminal1_batch),
}
res = TransitionData(obs_shape=self.obs_shape,
action_shape=self.action_shape)
for obs0, obs1, action, terminal, re in zip(result['obs0'],
result['obs1'],
result['actions'],
result['terminals1'],
result['rewards']):
res.append(state=obs0,
new_state=obs1,
action=action,
done=terminal,
reward=re)
return res
def test_trajectory_data(self):
env = make('Acrobot-v1')
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
a = TrajectoryData(env_spec)
tmp_traj = TransitionData(env_spec)
st = env.reset()
re_list = []
st_list = []
for i in range(100):
ac = env_spec.action_space.sample()
st_new, re, done, _ = env.step(action=ac)
st_list.append(st_new)
re_list.append(re)
if (i + 1) % 10 == 0:
done = True
else:
done = False
tmp_traj.append(state=st,
new_state=st_new,
action=ac,
done=done,
reward=re)
if done is True:
a.append(tmp_traj)
tmp_traj.reset()
self.assertEqual(a.trajectories.__len__(), 10)
for traj in a.trajectories:
self.assertEqual(len(traj), 10)
data = a.return_as_transition_data()
data_gen = data.return_generator()
for d, re, st in zip(data_gen, re_list, st_list):
self.assertEqual(d[3], re)
self.assertTrue(np.equal(st, d[1]).all())
def wrap_func():
mlp_dyna, local = self.create_continue_dynamics_model(
env_id='Pendulum-v0')
env_spec = local['env_spec']
env = local['env']
policy = func(env_spec=env_spec)[0]
algo, locals = self.create_mpc(env_spec=env_spec,
mlp_dyna=mlp_dyna,
policy=policy,
env=env)
algo.init()
for _ in range(100):
assert env_spec.action_space.contains(
algo.predict(env_spec.obs_space.sample()))
st = env.reset()
data = TransitionData(env_spec)
for _ in range(10):
ac = algo.predict(st)
new_st, re, done, _ = env.step(action=ac)
data.append(state=st,
new_state=new_st,
reward=re,
action=ac,
done=done)
print(algo.train(batch_data=data))
def test_StandScaler(self):
env = make('ModifiedHalfCheetah')
env_spec = env.env_spec
self.assertEqual(env_spec.flat_obs_dim, 18)
self.assertEqual(env_spec.flat_action_dim, 6)
buffer_size = 10
buffer = TransitionData(env_spec=env_spec, obs_shape=env_spec.obs_shape, \
action_shape=env_spec.action_shape, size=buffer_size)
obs = env.reset()
for i in range(buffer_size):
act = env.action_space.sample()
obs_, rew, done, _ = env.step(act)
buffer.append(obs, act, obs_, done, rew)
batch_list = buffer.sample_batch_as_Transition(4, all_as_batch=True)
state_input_scaler_1 = RunningStandardScaler(env_spec.flat_action_dim)
for batch_data in batch_list:
state_input_scaler_1.update_scaler(batch_data.action_set)
mean_1 = state_input_scaler_1._mean
var_1 = state_input_scaler_1._var
print(mean_1)
print(var_1)
state_input_scaler_2 = RunningStandardScaler(env_spec.flat_action_dim)
state_input_scaler_2.update_scaler(buffer.action_set)
mean_2 = state_input_scaler_2._mean
var_2 = state_input_scaler_2._var
print(mean_2)
print(var_2)
def test_apply_normalization(self):
'''
Test normalization & denormalization in Transition.apply_(de)normalization
'''
mlp_dyna, local = self.create_continue_dynamics_model(
env_id='ModifiedHalfCheetah', name='mlp_dyna_model')
mlp_dyna.init()
print(mlp_dyna.state_input_scaler)
env = local['env']
assert isinstance(env, ModifiedHalfCheetahEnv)
env_spec = env.env_spec
buffer_size = 50
random_buffer = TransitionData(env_spec=env_spec, obs_shape=env_spec.obs_shape, \
action_shape=env_spec.action_shape, size=buffer_size)
obs = env.reset()
for i in range(buffer_size):
act = env.action_space.sample()
obs_, reward, done, info = env.step(act)
random_buffer.append(obs, act, obs_, done, reward)
normalized_random_buffer, mean_dict, var_dict = random_buffer.apply_normalization(
)
denormalized_random_buffer = normalized_random_buffer.apply_denormalization(
None, mean_dict, var_dict)
self.assertEqual(random_buffer.action_set.any(),
denormalized_random_buffer.action_set.any())
self.assertEqual(random_buffer.state_set.any(),
denormalized_random_buffer.state_set.any())
def _launch(self) -> bool:
env = self.env
env_spec = self.env_spec
cyber = self.cyber
obs, ep_ret, ep_len = env.reset(), 0, 0
for step in range(self.total_steps):
self.step_counter.increase(1)
act = self.agent.predict(obs=obs)
obs_, reward, done, _ = cyber.step(obs, act)
_buffer = TransitionData(env_spec=env_spec, obs_shape=env_spec.obs_shape, action_shape=env_spec.action_shape)
_buffer.append(obs, act, obs_, done, reward)
self.agent.algo.append_to_memory(_buffer)
ep_ret += reward
ep_len += 1
if done or ep_len > self.max_step_per_episode:
obs, ep_ret, ep_len = env.reset(), 0, 0
else:
obs = obs_
if step > self.train_after_step and step % self.train_every_step == 0:
self.agent.train()
if step > self.test_after_step and step % self.test_every_step == 0:
self.data_sample, self.test_reward = self.agent.test(env=env,
cyber=cyber,
data_sample=self.data_sample,
test_reward=self.test_reward,
num_test=self.num_test,
max_step_per_episode=self.max_step_per_episode)
env.close()
self.plot_test_reward(self.data_sample, self.test_reward)
return True
def test_prior_eval(self):
env = make('Pendulum-v0')
name = 'demo_exp'
env_spec = EnvSpec(obs_space=env.observation_space, action_space=env.action_space)
data = TransitionData(env_spec=env_spec)
policy = UniformRandomPolicy(env_spec=env_spec)
# Do some initial sampling here to train gmm model
st = env.reset()
for i in range(100):
ac = policy.forward(st)
new_st, re, _, _ = env.step(ac)
data.append(state=st, new_state=new_st, action=ac, reward=re, done=False)
st = new_st
gmm = GaussianMixtureDynamicsPrior(env_spec=env_spec, batch_data=data)
gmm.init()
gmm.update(batch_data=data)
mu0, Phi, m, n0 = gmm.eval(batch_data=data)
state_shape = data.state_set.shape[1]
action_shape = data.action_set.shape[1]
self.assertEqual(state_shape + action_shape + state_shape, mu0.shape[0])
self.assertEqual(state_shape + action_shape + state_shape, Phi.shape[0])
self.assertEqual(state_shape + action_shape + state_shape, Phi.shape[1])
def test_Transition_union(self):
'''
Useless testcase.
'''
algo, locals = self.create_mpc(name='test_Transition_union')
env_spec = locals['env_spec']
env = locals['env']
env.env_spec = env_spec
algo.init()
for _ in range(100):
assert env_spec.action_space.contains(
algo.predict(env_spec.obs_space.sample()))
st = env.reset()
data = TransitionData(env_spec)
for _ in range(10):
ac = algo.predict(st)
new_st, re, done, _ = env.step(action=ac)
data.append(state=st,
new_state=new_st,
reward=re,
action=ac,
done=done)
print(algo.train(batch_data=data))
def test_init(self):
dqn, locals = self.create_dqn()
env = locals['env']
env_spec = locals['env_spec']
dqn.init()
st = env.reset()
a = TransitionData(env_spec)
for i in range(100):
ac = dqn.predict(obs=st, sess=self.sess, batch_flag=False)
st_new, re, done, _ = env.step(action=ac)
a.append(state=st, new_state=st_new, action=ac, done=done, reward=re)
st = st_new
dqn.append_to_memory(a)
new_dqn, _ = self.create_dqn(name='new_dqn')
new_dqn.copy_from(dqn)
self.assert_var_list_id_no_equal(dqn.q_value_func.parameters('tf_var_list'),
new_dqn.q_value_func.parameters('tf_var_list'))
self.assert_var_list_id_no_equal(dqn.target_q_value_func.parameters('tf_var_list'),
new_dqn.target_q_value_func.parameters('tf_var_list'))
self.assert_var_list_equal(dqn.q_value_func.parameters('tf_var_list'),
new_dqn.q_value_func.parameters('tf_var_list'))
self.assert_var_list_equal(dqn.target_q_value_func.parameters('tf_var_list'),
new_dqn.target_q_value_func.parameters('tf_var_list'))
dqn.save(save_path=GlobalConfig().DEFAULT_LOG_PATH + '/dqn_test',
global_step=0,
name=dqn.name)
for i in range(10):
print(dqn.train(batch_data=a, train_iter=10, sess=None, update_target=True))
print(dqn.train(batch_data=None, train_iter=10, sess=None, update_target=True))
self.assert_var_list_at_least_not_equal(dqn.q_value_func.parameters('tf_var_list'),
new_dqn.q_value_func.parameters('tf_var_list'))
self.assert_var_list_at_least_not_equal(dqn.target_q_value_func.parameters('tf_var_list'),
new_dqn.target_q_value_func.parameters('tf_var_list'))
dqn.load(path_to_model=GlobalConfig().DEFAULT_LOG_PATH + '/dqn_test',
model_name=dqn.name,
global_step=0)
self.assert_var_list_equal(dqn.q_value_func.parameters('tf_var_list'),
new_dqn.q_value_func.parameters('tf_var_list'))
self.assert_var_list_equal(dqn.target_q_value_func.parameters('tf_var_list'),
new_dqn.target_q_value_func.parameters('tf_var_list'))
for i in range(10):
self.sess.run(dqn.update_target_q_value_func_op,
feed_dict=dqn.parameters.return_tf_parameter_feed_dict())
var1 = self.sess.run(dqn.q_value_func.parameters('tf_var_list'))
var2 = self.sess.run(dqn.target_q_value_func.parameters('tf_var_list'))
import numpy as np
total_diff = 0.0
for v1, v2 in zip(var1, var2):
total_diff += np.mean(np.abs(np.array(v1) - np.array(v2)))
print('update target, difference mean', total_diff)
def get_some_samples(env, num, env_spec, policy):
data = TransitionData(env_spec=env_spec)
st = env.reset()
for i in range(num):
ac = policy.forward(st)
new_st, re, _, _ = env.step(ac)
data.append(state=st, new_state=new_st, action=ac, reward=re, done=False)
st = new_st
return data
def sample_transition(self, env, count=100):
data = TransitionData(env.env_spec)
st = env.get_state()
for i in range(count):
ac = env.env_spec.action_space.sample()
new_st, re, done, info = env.step(action=ac)
data.append(state=st,
action=ac,
new_state=new_st,
done=done,
reward=re)
return data
def test_mlp_dynamics_model(self):
mlp_dyna, local = self.create_continue_dynamics_model(
name='mlp_dyna_model')
env = local['env']
env_spec = local['env_spec']
env.reset()
mlp_dyna.init()
for i in range(100):
mlp_dyna.step(action=np.array(env_spec.action_space.sample()),
state=env_spec.obs_space.sample())
data = TransitionData(env_spec)
st = env.get_state()
for i in range(10):
ac = env_spec.action_space.sample()
new_st, re, done, info = env.step(action=ac)
data.append(state=st,
action=ac,
new_state=new_st,
done=done,
reward=re)
st = new_st
print(mlp_dyna.train(batch_data=data, train_iter=10))
mlp_dyna_2, _ = self.create_continue_dynamics_model(name='model_2')
mlp_dyna_2.init()
self.assert_var_list_at_least_not_equal(
var_list1=mlp_dyna.parameters('tf_var_list'),
var_list2=mlp_dyna_2.parameters('tf_var_list'))
self.assert_var_list_id_no_equal(
var_list1=mlp_dyna.parameters('tf_var_list'),
var_list2=mlp_dyna_2.parameters('tf_var_list'))
mlp_dyna_2.init(source_obj=mlp_dyna)
self.assert_var_list_equal(
var_list1=mlp_dyna.parameters('tf_var_list'),
var_list2=mlp_dyna_2.parameters('tf_var_list'))
self.assert_var_list_id_no_equal(
var_list1=mlp_dyna.parameters('tf_var_list'),
var_list2=mlp_dyna_2.parameters('tf_var_list'))
mlp_dyna_2.copy_from(mlp_dyna)
self.assert_var_list_equal(
var_list1=mlp_dyna.parameters('tf_var_list'),
var_list2=mlp_dyna_2.parameters('tf_var_list'))
self.assert_var_list_id_no_equal(
var_list1=mlp_dyna.parameters('tf_var_list'),
var_list2=mlp_dyna_2.parameters('tf_var_list'))
def test_dynamics_model_in_pendulum(self):
env = self.create_env('Pendulum-v0')
env_spec = EnvSpec(obs_space=env.observation_space, action_space=env.action_space)
policy, _ = self.create_uniform_policy(env_spec=env_spec)
data = TransitionData(env_spec=env_spec)
st = env.reset()
for i in range(100):
ac = policy.forward(st)
new_st, re, _, _ = env.step(ac)
data.append(state=st, new_state=new_st, action=ac, reward=re, done=False)
st = new_st
gp = GaussianProcessDyanmicsModel(env_spec=env_spec, batch_data=data)
gp.init()
gp.train()
for i in range(len(data.state_set)):
res = gp.step(action=data.action_set[i],
state=data.state_set[i],
allow_clip=True)
_, var = gp._state_transit(action=data.action_set[i],
state=data.state_set[i],
required_var=True)
print(res)
print(data.new_state_set[i])
print(np.sqrt(var))
# self.assertTrue(np.isclose(res,
# data.new_state_set[i], atol=1e-3).all())
self.assertTrue(np.greater(data.new_state_set[i] + 1.96 * np.sqrt(var), res).all())
self.assertTrue(np.less(data.new_state_set[i] - 1.96 * np.sqrt(var), res).all())
lengthscales = {}
variances = {}
noises = {}
for i, model in enumerate(gp.mgpr_model.models):
lengthscales['GP' + str(i)] = model.kern.lengthscales.value
variances['GP' + str(i)] = np.array([model.kern.variance.value])
noises['GP' + str(i)] = np.array([model.likelihood.variance.value])
print('-----Learned models------')
pd.set_option('precision', 3)
print('---Lengthscales---')
print(pd.DataFrame(data=lengthscales))
print('---Variances---')
print(pd.DataFrame(data=variances))
print('---Noises---')
print(pd.DataFrame(data=noises))
def test_dynamics_model_basic(self):
env = self.create_env('Pendulum-v0')
env_spec = EnvSpec(obs_space=env.observation_space, action_space=env.action_space)
policy, _ = self.create_uniform_policy(env_spec=env_spec)
data = TransitionData(env_spec=env_spec)
st = env.reset()
ac = policy.forward(st)
for i in range(10):
re = 0.0
data.append(state=np.ones_like(st) * 0.5, new_state=np.ones_like(st),
reward=re, done=False, action=np.ones_like(ac) * 0.1)
data.append(state=np.ones_like(st), new_state=np.ones_like(st) * 0.5,
reward=re, done=False, action=np.ones_like(ac) * -0.1)
gp = GaussianProcessDyanmicsModel(env_spec=env_spec, batch_data=data)
gp.init()
gp.train()
lengthscales = {}
variances = {}
noises = {}
i = 0
for model in gp.mgpr_model.models:
lengthscales['GP' + str(i)] = model.kern.lengthscales.value
variances['GP' + str(i)] = np.array([model.kern.variance.value])
noises['GP' + str(i)] = np.array([model.likelihood.variance.value])
i += 1
print('-----Learned models------')
pd.set_option('precision', 3)
print('---Lengthscales---')
print(pd.DataFrame(data=lengthscales))
print('---Variances---')
print(pd.DataFrame(data=variances))
print('---Noises---')
print(pd.DataFrame(data=noises))
for i in range(5):
self.assertTrue(np.isclose(gp.step(action=np.ones_like(ac) * -0.1,
state=np.ones_like(st)),
np.ones_like(st) * 0.5).all())
for i in range(5):
self.assertTrue(np.isclose(gp.step(action=np.ones_like(ac) * 0.1,
state=np.ones_like(st) * 0.5),
np.ones_like(st)).all())
for i in range(5):
print(gp.step(action=np.ones_like(ac) * -0.1,
state=np.ones_like(st) * 0.5))
def _sample_transitions(self, env: Env, agent, sample_count, init_state):
state = init_state
sample_record = TransitionData(env_spec=self.env_spec)
for i in range(sample_count):
action = agent.predict(obs=state)
new_state, re, done, info = env.step(action)
if not isinstance(done, bool):
raise TypeError()
sample_record.append(state=state,
action=action,
reward=re,
new_state=new_state,
done=done)
if done:
state = env.reset()
else:
state = new_state
return sample_record
def predict(self, obs, **kwargs):
if self.is_training is True:
return self.env_spec.action_space.sample()
rollout = TrajectoryData(env_spec=self.env_spec)
state = obs
for i in range(self.parameters('SAMPLED_PATH_NUM')):
path = TransitionData(env_spec=self.env_spec)
# todo terminal_func signal problem to be consider?
for _ in range(self.parameters('SAMPLED_HORIZON')):
ac = self.policy.forward(obs=state)
new_state, re, done, _ = self.dynamics_env.step(action=ac, state=state) # step() as an Env
path.append(state=state, action=ac, new_state=new_state, reward=re, done=done)
state = new_state
rollout.append(path)
rollout.trajectories.sort(key=lambda x: x.cumulative_reward, reverse=True)
ac = rollout.trajectories[0].action_set[0]
assert self.env_spec.action_space.contains(ac)
return ac
def test_random_buffer_1(self):
env = make('ModifiedHalfCheetah')
# env.init()
env_spec = env.env_spec
random_buffer = TransitionData(env_spec=env_spec, obs_shape=env_spec.obs_shape, \
action_shape=env_spec.action_shape, size=5)
rl_buffer = TransitionData(env_spec=env_spec, obs_shape=env_spec.obs_shape, \
action_shape=env_spec.action_shape, size=10)
max_step = 10
ep_len = 0
obs = env.reset()
while ep_len < max_step:
act = self.RandomController_get_action(env=env, state=obs)
obs_, reward, done, _ = env.step(act)
random_buffer.append(obs, act, obs_, done, reward)
assert not done
obs = obs_
ep_len += 1
def _sample_trajectories(self, env, agent, sample_count, init_state):
state = init_state
sample_record = TrajectoryData(self.env_spec)
done = False
for i in range(sample_count):
traj_record = TransitionData(self.env_spec)
while done is not True:
action = agent.predict(obs=state)
new_state, re, done, info = env.step(action)
if not isinstance(done, bool):
raise TypeError()
traj_record.append(state=state,
action=action,
reward=re,
new_state=new_state,
done=done)
state = new_state
state = env.reset()
sample_record.append(traj_record)
return sample_record
def test_init_continuous(self):
algo, locals = self.create_mpc(env_id='Pendulum-v0')
env_spec = locals['env_spec']
env = locals['env']
algo.init()
for _ in range(100):
assert env_spec.action_space.contains(
algo.predict(env_spec.obs_space.sample()))
st = env.reset()
data = TransitionData(env_spec)
for _ in range(10):
ac = algo.predict(st)
new_st, re, done, _ = env.step(action=ac)
data.append(state=st,
new_state=new_st,
reward=re,
action=ac,
done=done)
print(algo.train(batch_data=data))
def _test_1(self):
mlp_dyna, local = self.create_continue_dynamics_model(
env_id='ModifiedHalfCheetah', name='mlp_dyna_model')
print(local.items())
env = local['env']
assert isinstance(env, ModifiedHalfCheetahEnv)
env_spec = env.env_spec
batch_data = TransitionData(env_spec=env_spec, obs_shape=env_spec.obs_shape, \
action_shape=env_spec.action_shape)
batch_size = 32
obs = env.reset()
for i in range(batch_size):
act = self.RandomController_get_action(env=env, state=obs)
obs_, reward, done, info = env.step(action=act)
batch_data.append(obs, act, obs_, done, reward)
self.assertEqual(len(batch_data), batch_size)
mlp_dyna.init()
train_epoch = 20
for i in range(train_epoch):
res = mlp_dyna.train(batch_data, train_iter=10)
print('iter:{} loss:{}'.format(i, res))
def predict(self, obs, is_reward_func=True):
'''
Sample SAMPLED_PATH_NUM trajectories started from 'obs'. Return the optimal action.
:param obs: Initial state.
:param reverse_sort_flag: Decide the sort direction of trajectories, set to 'True' when using reward func.
:return: Optimal action for 'obs'.
'''
rollout = TrajectoryData(env_spec=self.env_spec)
for i in range(self.parameters('SAMPLED_PATH_NUM')):
path = TransitionData(env_spec=self.env_spec)
state = obs
for j in range(self.parameters('SAMPLED_HORIZON')):
act = self.policy.forward(obs=state) # env.action_space.sample()
new_state, cost, _, _ = self.dynamics_env.step(action=act, state=state) # step() as an Env
path.append(state=state, action=act, new_state=new_state, reward=cost, done=False)
state = new_state
rollout.append(path)
rollout.trajectories.sort(key=lambda x: x.cumulative_reward, reverse=is_reward_func)
optimal_act = rollout.trajectories[0].action_set[0]
assert self.env_spec.action_space.contains(optimal_act)
return optimal_act
from baconian.algo.dynamics.dynamics_model import DynamicsEnvWrapper
from baconian.algo.dynamics.terminal_func.terminal_func import RandomTerminalFunc
from baconian.algo.dynamics.reward_func.reward_func import RandomRewardFunc
env = make('Pendulum-v0')
name = 'demo_exp'
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
data = TransitionData(env_spec=env_spec)
policy = UniformRandomPolicy(env_spec=env_spec)
# Do some initial sampling here to train GP model
st = env.reset()
for i in range(100):
ac = policy.forward(st)
new_st, re, _, _ = env.step(ac)
data.append(state=st, new_state=new_st, action=ac, reward=re, done=False)
st = new_st
gp = GaussianProcessDyanmicsModel(env_spec=env_spec, batch_data=data)
gp.init()
gp.train()
dyna_env = DynamicsEnvWrapper(dynamics=gp)
# Since we only care about the prediction here, so we pass the terminal function and reward function setting with
# random one
dyna_env.set_terminal_reward_func(terminal_func=RandomTerminalFunc(),
reward_func=RandomRewardFunc())
st = env.reset()
real_state_list = []
dynamics_state_list = []
class PPO(ModelFreeAlgo, OnPolicyAlgo, MultiPlaceholderInput):
required_key_dict = DictConfig.load_json(
file_path=GlobalConfig().DEFAULT_PPO_REQUIRED_KEY_LIST)
@typechecked
def __init__(self,
env_spec: EnvSpec,
stochastic_policy: StochasticPolicy,
config_or_config_dict: (DictConfig, dict),
value_func: VValueFunction,
name='ppo'):
ModelFreeAlgo.__init__(self, env_spec=env_spec, name=name)
self.config = construct_dict_config(config_or_config_dict, self)
self.policy = stochastic_policy
self.value_func = value_func
to_ph_parameter_dict = dict()
self.trajectory_memory = TrajectoryData(env_spec=env_spec)
self.transition_data_for_trajectory = TransitionData(env_spec=env_spec)
self.value_func_train_data_buffer = None
# self.scaler = Scaler(obs_dim=self.env_spec.flat_obs_dim)
self.scaler = RunningStandardScaler(dims=self.env_spec.flat_obs_dim)
with tf.variable_scope(name):
self.advantages_ph = tf.placeholder(tf.float32, (None, ),
'advantages')
self.v_func_val_ph = tf.placeholder(tf.float32, (None, ),
'val_valfunc')
dist_info_list = self.policy.get_dist_info()
self.old_dist_tensor = [
(tf.placeholder(**dict(dtype=dist_info['dtype'],
shape=dist_info['shape'],
name=dist_info['name'])),
dist_info['name']) for dist_info in dist_info_list
]
self.old_policy = self.policy.make_copy(
reuse=False,
name_scope='old_{}'.format(self.policy.name),
name='old_{}'.format(self.policy.name),
distribution_tensors_tuple=tuple(self.old_dist_tensor))
to_ph_parameter_dict['beta'] = tf.placeholder(
tf.float32, (), 'beta')
to_ph_parameter_dict['eta'] = tf.placeholder(tf.float32, (), 'eta')
to_ph_parameter_dict['kl_target'] = tf.placeholder(
tf.float32, (), 'kl_target')
to_ph_parameter_dict['lr_multiplier'] = tf.placeholder(
tf.float32, (), 'lr_multiplier')
self.parameters = ParametersWithTensorflowVariable(
tf_var_list=[],
rest_parameters=dict(
advantages_ph=self.advantages_ph,
v_func_val_ph=self.v_func_val_ph,
),
to_ph_parameter_dict=to_ph_parameter_dict,
name='ppo_param',
save_rest_param_flag=False,
source_config=self.config,
require_snapshot=False)
with tf.variable_scope(name):
with tf.variable_scope('train'):
self.kl = tf.reduce_mean(self.old_policy.kl(other=self.policy))
self.policy_loss, self.policy_optimizer, self.policy_update_op = self._setup_policy_loss(
)
self.value_func_loss, self.value_func_optimizer, self.value_func_update_op = self._setup_value_func_loss(
)
var_list = get_tf_collection_var_list(
'{}/train'.format(name)) + self.policy_optimizer.variables(
) + self.value_func_optimizer.variables()
self.parameters.set_tf_var_list(
tf_var_list=sorted(list(set(var_list)), key=lambda x: x.name))
MultiPlaceholderInput.__init__(self,
sub_placeholder_input_list=[
dict(
obj=self.value_func,
attr_name='value_func',
),
dict(obj=self.policy,
attr_name='policy')
],
parameters=self.parameters)
@register_counter_info_to_status_decorator(increment=1,
info_key='init',
under_status='JUST_INITED')
def init(self, sess=None, source_obj=None):
self.policy.init()
self.value_func.init()
self.parameters.init()
if source_obj:
self.copy_from(source_obj)
super().init()
@record_return_decorator(which_recorder='self')
@register_counter_info_to_status_decorator(increment=1,
info_key='train',
under_status='TRAIN')
@typechecked
def train(self,
trajectory_data: TrajectoryData = None,
train_iter=None,
sess=None) -> dict:
super(PPO, self).train()
if trajectory_data is None:
trajectory_data = self.trajectory_memory
if len(trajectory_data) == 0:
raise MemoryBufferLessThanBatchSizeError(
'not enough trajectory data')
tf_sess = sess if sess else tf.get_default_session()
SampleProcessor.add_estimated_v_value(trajectory_data,
value_func=self.value_func)
SampleProcessor.add_discount_sum_reward(trajectory_data,
gamma=self.parameters('gamma'))
SampleProcessor.add_gae(trajectory_data,
gamma=self.parameters('gamma'),
name='advantage_set',
lam=self.parameters('lam'),
value_func=self.value_func)
train_data = trajectory_data.return_as_transition_data(
shuffle_flag=False)
SampleProcessor.normalization(train_data, key='advantage_set')
policy_res_dict = self._update_policy(
train_data=train_data,
train_iter=train_iter
if train_iter else self.parameters('policy_train_iter'),
sess=tf_sess)
value_func_res_dict = self._update_value_func(
train_data=train_data,
train_iter=train_iter
if train_iter else self.parameters('value_func_train_iter'),
sess=tf_sess)
self.trajectory_memory.reset()
return {**policy_res_dict, **value_func_res_dict}
@register_counter_info_to_status_decorator(increment=1,
info_key='test',
under_status='TEST')
def test(self, *arg, **kwargs) -> dict:
return super().test(*arg, **kwargs)
@register_counter_info_to_status_decorator(increment=1, info_key='predict')
@typechecked
def predict(self, obs: np.ndarray, sess=None, batch_flag: bool = False):
tf_sess = sess if sess else tf.get_default_session()
obs = make_batch(obs, original_shape=self.env_spec.obs_shape)
obs = self.scaler.process(data=obs)
ac = self.policy.forward(
obs=obs,
sess=tf_sess,
feed_dict=self.parameters.return_tf_parameter_feed_dict())
return ac
@typechecked
def append_to_memory(self, samples: SampleData):
# todo how to make sure the data's time sequential
iter_samples = samples.return_generator()
# scale, offset = self.scaler.get()
obs_list = []
for state, new_state, action, reward, done in iter_samples:
obs_list.append(state)
self.transition_data_for_trajectory.append(
state=self.scaler.process(state),
new_state=self.scaler.process(new_state),
action=action,
reward=reward,
done=done)
if done is True:
self.trajectory_memory.append(
self.transition_data_for_trajectory)
self.transition_data_for_trajectory.reset()
self.scaler.update_scaler(data=np.array(obs_list))
@record_return_decorator(which_recorder='self')
def save(self, global_step, save_path=None, name=None, **kwargs):
save_path = save_path if save_path else GlobalConfig(
).DEFAULT_MODEL_CHECKPOINT_PATH
name = name if name else self.name
MultiPlaceholderInput.save(self,
save_path=save_path,
global_step=global_step,
name=name,
**kwargs)
return dict(check_point_save_path=save_path,
check_point_save_global_step=global_step,
check_point_save_name=name)
@record_return_decorator(which_recorder='self')
def load(self, path_to_model, model_name, global_step=None, **kwargs):
MultiPlaceholderInput.load(self, path_to_model, model_name,
global_step, **kwargs)
return dict(check_point_load_path=path_to_model,
check_point_load_global_step=global_step,
check_point_load_name=model_name)
def _setup_policy_loss(self):
"""
Code clip from pat-cody
Three loss terms:
1) standard policy gradient
2) D_KL(pi_old || pi_new)
3) Hinge loss on [D_KL - kl_targ]^2
See: https://arxiv.org/pdf/1707.02286.pdf
"""
if self.parameters('clipping_range') is not None:
pg_ratio = tf.exp(self.policy.log_prob() -
self.old_policy.log_prob())
clipped_pg_ratio = tf.clip_by_value(
pg_ratio, 1 - self.parameters('clipping_range')[0],
1 + self.parameters('clipping_range')[1])
surrogate_loss = tf.minimum(self.advantages_ph * pg_ratio,
self.advantages_ph * clipped_pg_ratio)
loss = -tf.reduce_mean(surrogate_loss)
else:
loss1 = -tf.reduce_mean(
self.advantages_ph *
tf.exp(self.policy.log_prob() - self.old_policy.log_prob()))
loss2 = tf.reduce_mean(self.parameters('beta') * self.kl)
loss3 = self.parameters('eta') * tf.square(
tf.maximum(0.0, self.kl - 2.0 * self.parameters('kl_target')))
loss = loss1 + loss2 + loss3
self.loss1 = loss1
self.loss2 = loss2
self.loss3 = loss3
if isinstance(self.policy, PlaceholderInput):
reg_list = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
scope=self.policy.name_scope)
if len(reg_list) > 0:
reg_loss = tf.reduce_sum(reg_list)
loss += reg_loss
optimizer = tf.train.AdamOptimizer(
learning_rate=self.parameters('policy_lr') *
self.parameters('lr_multiplier'))
train_op = optimizer.minimize(
loss, var_list=self.policy.parameters('tf_var_list'))
return loss, optimizer, train_op
def _setup_value_func_loss(self):
# todo update the value_func design
loss = tf.reduce_mean(
tf.square(self.value_func.v_tensor - self.v_func_val_ph))
reg_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
scope=self.value_func.name_scope)
if len(reg_loss) > 0:
loss += tf.reduce_sum(reg_loss)
optimizer = tf.train.AdamOptimizer(self.parameters('value_func_lr'))
train_op = optimizer.minimize(
loss, var_list=self.value_func.parameters('tf_var_list'))
return loss, optimizer, train_op
def _update_policy(self, train_data: TransitionData, train_iter, sess):
old_policy_feed_dict = dict()
res = sess.run(
[
getattr(self.policy, tensor[1])
for tensor in self.old_dist_tensor
],
feed_dict={
self.policy.parameters('state_input'): train_data('state_set'),
self.policy.parameters('action_input'):
train_data('action_set'),
**self.parameters.return_tf_parameter_feed_dict()
})
for tensor, val in zip(self.old_dist_tensor, res):
old_policy_feed_dict[tensor[0]] = val
feed_dict = {
self.policy.parameters('action_input'): train_data('action_set'),
self.old_policy.parameters('action_input'):
train_data('action_set'),
self.policy.parameters('state_input'): train_data('state_set'),
self.advantages_ph: train_data('advantage_set'),
**self.parameters.return_tf_parameter_feed_dict(),
**old_policy_feed_dict
}
average_loss, average_kl, average_entropy = 0.0, 0.0, 0.0
total_epoch = 0
kl = None
for i in range(train_iter):
loss, kl, entropy, _ = sess.run([
self.policy_loss, self.kl,
tf.reduce_mean(self.policy.entropy()), self.policy_update_op
],
feed_dict=feed_dict)
average_loss += loss
average_kl += kl
average_entropy += entropy
total_epoch = i + 1
if kl > self.parameters('kl_target', require_true_value=True) * 4:
# early stopping if D_KL diverges badly
break
average_loss, average_kl, average_entropy = average_loss / total_epoch, average_kl / total_epoch, average_entropy / total_epoch
if kl > self.parameters('kl_target', require_true_value=True
) * 2: # servo beta to reach D_KL target
self.parameters.set(
key='beta',
new_val=np.minimum(
35,
1.5 * self.parameters('beta', require_true_value=True)))
if self.parameters(
'beta', require_true_value=True) > 30 and self.parameters(
'lr_multiplier', require_true_value=True) > 0.1:
self.parameters.set(
key='lr_multiplier',
new_val=self.parameters('lr_multiplier',
require_true_value=True) / 1.5)
elif kl < self.parameters('kl_target', require_true_value=True) / 2:
self.parameters.set(
key='beta',
new_val=np.maximum(
1 / 35,
self.parameters('beta', require_true_value=True) / 1.5))
if self.parameters('beta', require_true_value=True) < (
1 / 30) and self.parameters('lr_multiplier',
require_true_value=True) < 10:
self.parameters.set(
key='lr_multiplier',
new_val=self.parameters('lr_multiplier',
require_true_value=True) * 1.5)
return dict(policy_average_loss=average_loss,
policy_average_kl=average_kl,
policy_average_entropy=average_entropy,
policy_total_train_epoch=total_epoch)
def _update_value_func(self, train_data: TransitionData, train_iter, sess):
if self.value_func_train_data_buffer is None:
self.value_func_train_data_buffer = train_data
else:
self.value_func_train_data_buffer.union(train_data)
y_hat = self.value_func.forward(obs=train_data('state_set'))
old_exp_var = 1 - np.var(train_data('advantage_set') - y_hat) / np.var(
train_data('advantage_set'))
for i in range(train_iter):
data_gen = self.value_func_train_data_buffer.return_generator(
batch_size=self.parameters('value_func_train_batch_size'),
infinite_run=False,
shuffle_flag=True,
assigned_keys=('state_set', 'new_state_set', 'action_set',
'reward_set', 'done_set', 'advantage_set'))
for batch in data_gen:
loss, _ = sess.run(
[self.value_func_loss, self.value_func_update_op],
feed_dict={
self.value_func.state_input: batch[0],
self.v_func_val_ph: batch[5],
**self.parameters.return_tf_parameter_feed_dict()
})
y_hat = self.value_func.forward(obs=train_data('state_set'))
loss = np.mean(np.square(y_hat - train_data('advantage_set')))
exp_var = 1 - np.var(train_data('advantage_set') - y_hat) / np.var(
train_data('advantage_set'))
self.value_func_train_data_buffer = train_data
return dict(value_func_loss=loss,
value_func_policy_exp_var=exp_var,
value_func_policy_old_exp_var=old_exp_var)
def test_l1_l2_norm(self):
env = make('Acrobot-v1')
env_spec = EnvSpec(obs_space=env.observation_space,
action_space=env.action_space)
name = 'dqn'
mlp_q = MLPQValueFunction(env_spec=env_spec,
name_scope=name + '_mlp',
name=name + '_mlp',
mlp_config=[{
"ACT": "RELU",
"B_INIT_VALUE": 0.0,
"NAME": "1",
"N_UNITS": 16,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03,
"L1_NORM": 1000.0,
"L2_NORM": 1000.0
}, {
"ACT": "LINEAR",
"B_INIT_VALUE": 0.0,
"NAME": "OUPTUT",
"N_UNITS": 1,
"L1_NORM": 1000.0,
"L2_NORM": 1000.0,
"TYPE": "DENSE",
"W_NORMAL_STDDEV": 0.03
}])
dqn = DQN(env_spec=env_spec,
config_or_config_dict=dict(REPLAY_BUFFER_SIZE=1000,
GAMMA=0.99,
BATCH_SIZE=10,
LEARNING_RATE=0.01,
TRAIN_ITERATION=1,
DECAY=0.5),
name=name,
value_func=mlp_q)
dqn2, _ = self.create_dqn(name='dqn_2')
a = TransitionData(env_spec)
st = env.reset()
dqn.init()
dqn2.init()
for i in range(100):
ac = dqn.predict(obs=st, sess=self.sess, batch_flag=False)
st_new, re, done, _ = env.step(action=ac)
a.append(state=st,
new_state=st_new,
action=ac,
done=done,
reward=re)
st = st_new
dqn.append_to_memory(a)
for i in range(20):
print(
'dqn1 loss: ',
dqn.train(batch_data=a,
train_iter=10,
sess=None,
update_target=True))
print(
'dqn2 loss: ',
dqn2.train(batch_data=a,
train_iter=10,
sess=None,
update_target=True))
var_list = self.sess.run(dqn.q_value_func.parameters('tf_var_list'))
print(var_list)
var_list2 = self.sess.run(dqn2.q_value_func.parameters('tf_var_list'))
print(var_list2)
for var, var2 in zip(var_list, var_list2):
diff = np.abs(var2) - np.abs(var)
self.assertTrue(np.greater(np.mean(diff), 0.0).all())
def test_init(self):
ppo, locals = self.create_ppo()
env = locals['env']
env_spec = locals['env_spec']
ppo.init()
new_ppo, _ = self.create_ppo(name='new_ppo')
new_ppo.copy_from(ppo)
self.assert_var_list_id_no_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_id_no_equal(
ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))
self.assert_var_list_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_equal(ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))
data = TransitionData(env_spec)
st = env.reset()
for i in range(100):
ac = ppo.predict(st)
assert ac.shape[0] == 1
self.assertTrue(env_spec.action_space.contains(ac[0]))
new_st, re, done, _ = env.step(ac)
if i % 9 == 0 and i > 0:
done = True
else:
done = False
data.append(state=st,
new_state=new_st,
action=ac,
reward=re,
done=done)
traj = TrajectoryData(env_spec=env_spec)
traj.append(data)
ppo.append_to_memory(traj)
ppo.save(save_path=GlobalConfig().DEFAULT_LOG_PATH + '/ppo_test',
global_step=0,
name=ppo.name)
for i in range(5):
ppo.append_to_memory(traj)
res = ppo.train()
print('value_func_loss {}, policy_average_loss: {}'.format(
res['value_func_loss'], res['policy_average_loss']))
traj_data = TrajectoryData(env_spec=env_spec)
traj_data.append(data)
res = ppo.train(trajectory_data=traj_data,
train_iter=5,
sess=self.sess)
print('value_func_loss {}, policy_average_loss: {}'.format(
res['value_func_loss'], res['policy_average_loss']))
self.assert_var_list_at_least_not_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_at_least_not_equal(
ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))
ppo.load(path_to_model=GlobalConfig().DEFAULT_LOG_PATH + '/ppo_test',
model_name=ppo.name,
global_step=0)
self.assert_var_list_equal(
ppo.value_func.parameters('tf_var_list'),
new_ppo.value_func.parameters('tf_var_list'))
self.assert_var_list_equal(ppo.policy.parameters('tf_var_list'),
new_ppo.policy.parameters('tf_var_list'))