def __init__(
self,
name,
env_spec,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
hidden_sizes=[],
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax,
prob_network=None,
):
"""
:param env_spec: A spec for the mdp.
:param hidden_sizes: list of sizes for the fully connected hidden layers
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param prob_network: manually specified network for this policy, other network params
are ignored
:return:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Discrete)
self._env_spec = env_spec
# import pdb; pdb.set_trace()
if prob_network is None:
prob_network = ConvNetwork(
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.n,
conv_filters=conv_filters,
conv_filter_sizes=conv_filter_sizes,
conv_strides=conv_strides,
conv_pads=conv_pads,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name="prob_network",
)
self._l_prob = prob_network.output_layer
self._l_obs = prob_network.input_layer
self._f_prob = tensor_utils.compile_function(
[prob_network.input_layer.input_var],
L.get_output(prob_network.output_layer))
self._dist = Categorical(env_spec.action_space.n)
super(CategoricalConvPolicy, self).__init__(env_spec)
LayersPowered.__init__(self, [prob_network.output_layer])
def get_value_network(env):
value_network = ConvNetwork(
name='value_network',
input_shape=env.observation_space.shape,
output_dim=1,
# number of channels/filters for each conv layer
conv_filters=(16, 32),
# filter size
conv_filter_sizes=(8, 4),
conv_strides=(4, 2),
conv_pads=('VALID', 'VALID'),
hidden_sizes=(256, ),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
batch_normalization=False)
return value_network
def get_policy_network(env):
policy_network = ConvNetwork(
name='prob_network',
input_shape=env.observation_space.shape,
output_dim=env.action_space.n,
# number of channels/filters for each conv layer
conv_filters=(16, 32),
# filter size
conv_filter_sizes=(8, 4),
conv_strides=(4, 2),
conv_pads=('VALID', 'VALID'),
hidden_sizes=(256, ),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax,
batch_normalization=False,
)
return policy_network
def _make_subnetwork(self,
input_layer,
dim_output,
hidden_sizes,
output_nonlinearity=tf.sigmoid,
hidden_nonlinearity=tf.nn.tanh,
name="pred_network",
conv_filters=None,
conv_filter_sizes=None,
conv_strides=None,
conv_pads=None,
input_shape=None):
if conv_filters is not None:
input_layer = L.reshape(input_layer, ([0], ) + input_shape,
name="reshape_input")
prob_network = ConvNetwork(input_shape=input_shape,
output_dim=dim_output,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name=name,
input_layer=input_layer,
conv_filters=conv_filters,
conv_filter_sizes=conv_filter_sizes,
conv_strides=conv_strides,
conv_pads=conv_pads)
else:
prob_network = MLP(output_dim=dim_output,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
name=name,
input_layer=input_layer)
return prob_network.output_layer
stub(globals())
# Params range
seeds = range(0, 5)
for seed in seeds:
env = TfEnv(normalize(env=GymEnv('Box3dReachPixel-v14',record_video=False, \
log_dir='/tmp/gym_test',record_log=False)))
env_spec = env.spec
policy_cnn = ConvNetwork(
name="policy_conv_network",
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.flat_dim,
conv_filters=(64, 64, 64, 32),
conv_filter_sizes=((5, 5), (3, 3), (3, 3), (3, 3)),
conv_strides=(3, 3, 3, 2),
conv_pads=('SAME', 'SAME', 'SAME', 'SAME'),
hidden_sizes=(256, ),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
)
baseline_cnn = ConvNetwork(
name="baseline_conv_network",
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.flat_dim,
conv_filters=(64, 64, 64, 32),
conv_filter_sizes=((5, 5), (3, 3), (3, 3), (3, 3)),
conv_strides=(3, 3, 3, 2),
conv_pads=('SAME', 'SAME', 'SAME', 'SAME'),
hidden_sizes=(256, ),
stub(globals())
# Params range
seeds = range(2, 4)
for seed in seeds:
env = TfEnv(normalize(env=GymEnv('Box3dReachPixel-v6',record_video=False, \
log_dir='/tmp/gym_test',record_log=False)))
env_spec = env.spec
cnn = ConvNetwork(
name="conv_feature_network",
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.flat_dim,
conv_filters=(32, 32, 32, 32, 32),
conv_filter_sizes=((3,3),(3,3),(3,3),(3,3), (3,3)),
conv_strides=(2, 2, 2, 2, 2),
conv_pads=('SAME', 'SAME', 'SAME', 'SAME', 'SAME'),
hidden_sizes=(256,),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
)
policy = GaussianConvFeaturePolicy(
"conv_feature_policy",
env_spec=env_spec,
feature_network=cnn,
hidden_sizes=(128,64),
output_nonlinearity=tf.nn.tanh,
)
baseline = NNBaseline(
def forward_CNN_MLP(
self,
name,
all_params,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
conv_output_dim,
conv_hidden_sizes, # new
input_tensor=None,
batch_normalization=False,
reuse=True,
is_training=False):
# is_training and reuse are for batch norm, irrelevant if batch_norm set to False
# set reuse to False if the first time this func is called.
with tf.variable_scope(name):
if input_tensor is None:
l_in = make_input(shape=self.input_total_shape,
input_var=None,
name='input')
else:
l_in = input_tensor
# l_img_in = l_in[:][:np.prod(self.input_img_shape)]
l_img_in = tf.slice(l_in, [0, 0],
[-1, np.prod(self.input_img_shape)])
# l_state_in = l_in[:][np.prod(self.input_img_shape):]
l_state_in = tf.slice(l_in, [0, np.prod(self.input_img_shape)],
[-1, -1])
# print("Debug212",l_img_in, l_state_in)
l_normalized_img_in = tf.cast(l_img_in, tf.float32) / 255
# if self.cnn is None:
self.cnn = ConvNetwork(
name=name + "cnn",
input_shape=self.input_img_shape,
output_dim=conv_output_dim,
conv_filters=conv_filters,
conv_filter_sizes=conv_filter_sizes,
conv_strides=conv_strides,
conv_pads=conv_pads,
hidden_sizes=conv_hidden_sizes,
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=L.spatial_expected_softmax,
input_var=l_normalized_img_in)
# cnn = self.cnn
# else:
# self.cnn2 = ConvNetwork(name=name+"cnn2",input_shape=self.input_img_shape,output_dim=conv_output_dim,
# conv_filters=conv_filters,conv_filter_sizes=conv_filter_sizes,conv_strides=conv_strides,
# conv_pads=conv_pads, hidden_sizes=conv_hidden_sizes,hidden_nonlinearity=tf.nn.relu,output_nonlinearity=L.spatial_expected_softmax, input_var=l_normalized_img_in)
# cnn = self.cnn2
# print("debug234, cnn output layer", L.get_output(self.cnn._l_out))
l_hid = tf.concat(
[l_state_in, L.get_output(self.cnn.output_layer)], -1,
'post_conv_input')
# l_hid = tf.concat([l_state_in,l_normalized_img_in],-1,'post_conv_input')
# l_hid=l_in
for idx in range(self.n_hidden):
l_hid = forward_dense_layer(
l_hid,
all_params['W' + str(idx)],
all_params['b' + str(idx)],
batch_norm=batch_normalization,
nonlinearity=self.hidden_nonlinearity,
scope=str(idx),
reuse=reuse,
is_training=is_training)
output = forward_dense_layer(
l_hid,
all_params['W' + str(self.n_hidden)],
all_params['b' + str(self.n_hidden)],
batch_norm=False,
nonlinearity=self.output_nonlinearity,
)
return l_in, output
def rllab_envpolicy_parser(env, args):
if isinstance(args, dict):
args = tonamedtuple(args)
env = RLLabEnv(env, mode=args.control)
if args.algo[:2] == 'tf':
env = TfEnv(env)
# Policy
if args.recurrent:
if args.feature_net:
feature_network = MLP(
name='feature_net',
input_shape=(env.spec.observation_space.flat_dim +
env.spec.action_space.flat_dim, ),
output_dim=args.feature_output,
hidden_sizes=tuple(args.feature_hidden),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None)
elif args.conv:
strides = tuple(args.conv_strides)
chans = tuple(args.conv_channels)
filts = tuple(args.conv_filters)
assert len(strides) == len(chans) == len(
filts), "strides, chans and filts not equal"
# only discrete actions supported, should be straightforward to extend to continuous
assert isinstance(
env.spec.action_space,
Discrete), "Only discrete action spaces support conv"
feature_network = ConvNetwork(
name='feature_net',
input_shape=env.spec.observation_space.shape,
output_dim=args.feature_output,
conv_filters=chans,
conv_filter_sizes=filts,
conv_strides=strides,
conv_pads=('VALID', ) * len(chans),
hidden_sizes=tuple(args.feature_hidden),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None)
else:
feature_network = None
if args.recurrent == 'gru':
if isinstance(env.spec.action_space, Box):
policy = GaussianGRUPolicy(env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(
args.policy_hidden[0]),
name='policy')
elif isinstance(env.spec.action_space, Discrete):
policy = CategoricalGRUPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden[0]),
name='policy',
state_include_action=False if args.conv else True)
else:
raise NotImplementedError(env.spec.observation_space)
elif args.recurrent == 'lstm':
if isinstance(env.spec.action_space, Box):
policy = GaussianLSTMPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden),
name='policy')
elif isinstance(env.spec.action_space, Discrete):
policy = CategoricalLSTMPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden),
name='policy')
else:
raise NotImplementedError(env.spec.action_space)
else:
raise NotImplementedError(args.recurrent)
elif args.conv:
strides = tuple(args.conv_strides)
chans = tuple(args.conv_channels)
filts = tuple(args.conv_filters)
assert len(strides) == len(chans) == len(
filts), "strides, chans and filts not equal"
# only discrete actions supported, should be straightforward to extend to continuous
assert isinstance(
env.spec.action_space,
Discrete), "Only discrete action spaces support conv"
feature_network = ConvNetwork(
name='feature_net',
input_shape=env.spec.observation_space.shape,
output_dim=env.spec.action_space.n,
conv_filters=chans,
conv_filter_sizes=filts,
conv_strides=strides,
conv_pads=('VALID', ) * len(chans),
hidden_sizes=tuple(args.policy_hidden),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax)
policy = CategoricalMLPPolicy(name='policy',
env_spec=env.spec,
prob_network=feature_network)
else:
if isinstance(env.spec.action_space, Box):
policy = GaussianMLPPolicy(env_spec=env.spec,
hidden_sizes=tuple(
args.policy_hidden),
min_std=args.min_std,
name='policy')
elif isinstance(env.spec.action_space, Discrete):
policy = CategoricalMLPPolicy(env_spec=env.spec,
hidden_sizes=tuple(
args.policy_hidden),
name='policy')
else:
raise NotImplementedError(env.spec.action_space)
elif args.algo[:2] == 'th':
# Policy
if args.recurrent:
if args.feature_net:
feature_network = thMLP(
input_shape=(env.spec.observation_space.flat_dim +
env.spec.action_space.flat_dim, ),
output_dim=args.feature_output,
hidden_sizes=tuple(args.feature_hidden),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None)
else:
feature_network = None
if args.recurrent == 'gru':
if isinstance(env.spec.observation_space, thBox):
policy = thGaussianGRUPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden[0]),
)
elif isinstance(env.spec.observation_space, thDiscrete):
policy = thCategoricalGRUPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden[0]),
)
else:
raise NotImplementedError(env.spec.observation_space)
# elif args.recurrent == 'lstm':
# if isinstance(env.spec.action_space, thBox):
# policy = thGaussianLSTMPolicy(env_spec=env.spec,
# feature_network=feature_network,
# hidden_dim=int(args.policy_hidden),
# name='policy')
# elif isinstance(env.spec.action_space, thDiscrete):
# policy = thCategoricalLSTMPolicy(env_spec=env.spec,
# feature_network=feature_network,
# hidden_dim=int(args.policy_hidden),
# name='policy')
# else:
# raise NotImplementedError(env.spec.action_space)
else:
raise NotImplementedError(args.recurrent)
else:
if args.algo == 'thddpg':
assert isinstance(env.spec.action_space, thBox)
policy = thDeterministicMLPPolicy(
env_spec=env.spec,
hidden_sizes=tuple(args.policy_hidden),
)
else:
if isinstance(env.spec.action_space, thBox):
policy = thGaussianMLPPolicy(env_spec=env.spec,
hidden_sizes=tuple(
args.policy_hidden),
min_std=args.min_std)
elif isinstance(env.spec.action_space, thDiscrete):
policy = thCategoricalMLPPolicy(env_spec=env.spec,
hidden_sizes=tuple(
args.policy_hidden),
min_std=args.min_std)
else:
raise NotImplementedError(env.spec.action_space)
if args.control == 'concurrent':
return env, policies
else:
return env, policy
def main():
now = datetime.datetime.now(dateutil.tz.tzlocal())
rand_id = str(uuid.uuid4())[:5]
timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z')
default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id)
parser = argparse.ArgumentParser()
parser.add_argument('--exp_name',
type=str,
default=default_exp_name,
help='Name of the experiment.')
parser.add_argument('--discount', type=float, default=0.99)
parser.add_argument('--gae_lambda', type=float, default=1.0)
parser.add_argument('--reward_scale', type=float, default=1.0)
parser.add_argument('--n_iter', type=int, default=250)
parser.add_argument('--sampler_workers', type=int, default=1)
parser.add_argument('--max_traj_len', type=int, default=250)
parser.add_argument('--update_curriculum',
action='store_true',
default=False)
parser.add_argument('--n_timesteps', type=int, default=8000)
parser.add_argument('--control', type=str, default='centralized')
parser.add_argument('--rectangle', type=str, default='10,10')
parser.add_argument('--map_type', type=str, default='rectangle')
parser.add_argument('--n_evaders', type=int, default=5)
parser.add_argument('--n_pursuers', type=int, default=2)
parser.add_argument('--obs_range', type=int, default=3)
parser.add_argument('--n_catch', type=int, default=2)
parser.add_argument('--urgency', type=float, default=0.0)
parser.add_argument('--pursuit', dest='train_pursuit', action='store_true')
parser.add_argument('--evade', dest='train_pursuit', action='store_false')
parser.set_defaults(train_pursuit=True)
parser.add_argument('--surround', action='store_true', default=False)
parser.add_argument('--constraint_window', type=float, default=1.0)
parser.add_argument('--sample_maps', action='store_true', default=False)
parser.add_argument('--map_file', type=str, default='../maps/map_pool.npy')
parser.add_argument('--flatten', action='store_true', default=False)
parser.add_argument('--reward_mech', type=str, default='global')
parser.add_argument('--catchr', type=float, default=0.1)
parser.add_argument('--term_pursuit', type=float, default=5.0)
parser.add_argument('--recurrent', type=str, default=None)
parser.add_argument('--policy_hidden_sizes', type=str, default='128,128')
parser.add_argument('--baselin_hidden_sizes', type=str, default='128,128')
parser.add_argument('--baseline_type', type=str, default='linear')
parser.add_argument('--conv', action='store_true', default=False)
parser.add_argument('--max_kl', type=float, default=0.01)
parser.add_argument('--checkpoint', type=str, default=None)
parser.add_argument('--log_dir', type=str, required=False)
parser.add_argument('--tabular_log_file',
type=str,
default='progress.csv',
help='Name of the tabular log file (in csv).')
parser.add_argument('--text_log_file',
type=str,
default='debug.log',
help='Name of the text log file (in pure text).')
parser.add_argument('--params_log_file',
type=str,
default='params.json',
help='Name of the parameter log file (in json).')
parser.add_argument('--seed', type=int, help='Random seed for numpy')
parser.add_argument('--args_data',
type=str,
help='Pickled data for stub objects')
parser.add_argument('--snapshot_mode',
type=str,
default='all',
help='Mode to save the snapshot. Can be either "all" '
'(all iterations will be saved), "last" (only '
'the last iteration will be saved), or "none" '
'(do not save snapshots)')
parser.add_argument(
'--log_tabular_only',
type=ast.literal_eval,
default=False,
help=
'Whether to only print the tabular log information (in a horizontal format)'
)
args = parser.parse_args()
parallel_sampler.initialize(n_parallel=args.sampler_workers)
if args.seed is not None:
set_seed(args.seed)
parallel_sampler.set_seed(args.seed)
args.hidden_sizes = tuple(map(int, args.policy_hidden_sizes.split(',')))
if args.checkpoint:
with tf.Session() as sess:
data = joblib.load(args.checkpoint)
policy = data['policy']
env = data['env']
else:
if args.sample_maps:
map_pool = np.load(args.map_file)
else:
if args.map_type == 'rectangle':
env_map = TwoDMaps.rectangle_map(
*map(int, args.rectangle.split(',')))
elif args.map_type == 'complex':
env_map = TwoDMaps.complex_map(
*map(int, args.rectangle.split(',')))
else:
raise NotImplementedError()
map_pool = [env_map]
env = PursuitEvade(map_pool,
n_evaders=args.n_evaders,
n_pursuers=args.n_pursuers,
obs_range=args.obs_range,
n_catch=args.n_catch,
train_pursuit=args.train_pursuit,
urgency_reward=args.urgency,
surround=args.surround,
sample_maps=args.sample_maps,
constraint_window=args.constraint_window,
flatten=args.flatten,
reward_mech=args.reward_mech,
catchr=args.catchr,
term_pursuit=args.term_pursuit)
env = TfEnv(
RLLabEnv(StandardizedEnv(env,
scale_reward=args.reward_scale,
enable_obsnorm=False),
mode=args.control))
if args.recurrent:
if args.conv:
feature_network = ConvNetwork(
name='feature_net',
input_shape=emv.spec.observation_space.shape,
output_dim=5,
conv_filters=(16, 32, 32),
conv_filter_sizes=(3, 3, 3),
conv_strides=(1, 1, 1),
conv_pads=('VALID', 'VALID', 'VALID'),
hidden_sizes=(64, ),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax)
else:
feature_network = MLP(
name='feature_net',
input_shape=(env.spec.observation_space.flat_dim +
env.spec.action_space.flat_dim, ),
output_dim=5,
hidden_sizes=(256, 128, 64),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None)
if args.recurrent == 'gru':
policy = CategoricalGRUPolicy(env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(
args.policy_hidden_sizes),
name='policy')
elif args.recurrent == 'lstm':
policy = CategoricalLSTMPolicy(env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(
args.policy_hidden_sizes),
name='policy')
elif args.conv:
feature_network = ConvNetwork(
name='feature_net',
input_shape=env.spec.observation_space.shape,
output_dim=5,
conv_filters=(8, 16),
conv_filter_sizes=(3, 3),
conv_strides=(2, 1),
conv_pads=('VALID', 'VALID'),
hidden_sizes=(32, ),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax)
policy = CategoricalMLPPolicy(name='policy',
env_spec=env.spec,
prob_network=feature_network)
else:
policy = CategoricalMLPPolicy(name='policy',
env_spec=env.spec,
hidden_sizes=args.hidden_sizes)
if args.baseline_type == 'linear':
baseline = LinearFeatureBaseline(env_spec=env.spec)
else:
baseline = ZeroBaseline(env_spec=env.spec)
# logger
default_log_dir = config.LOG_DIR
if args.log_dir is None:
log_dir = osp.join(default_log_dir, args.exp_name)
else:
log_dir = args.log_dir
tabular_log_file = osp.join(log_dir, args.tabular_log_file)
text_log_file = osp.join(log_dir, args.text_log_file)
params_log_file = osp.join(log_dir, args.params_log_file)
logger.log_parameters_lite(params_log_file, args)
logger.add_text_output(text_log_file)
logger.add_tabular_output(tabular_log_file)
prev_snapshot_dir = logger.get_snapshot_dir()
prev_mode = logger.get_snapshot_mode()
logger.set_snapshot_dir(log_dir)
logger.set_snapshot_mode(args.snapshot_mode)
logger.set_log_tabular_only(args.log_tabular_only)
logger.push_prefix("[%s] " % args.exp_name)
algo = TRPO(
env=env,
policy=policy,
baseline=baseline,
batch_size=args.n_timesteps,
max_path_length=args.max_traj_len,
n_itr=args.n_iter,
discount=args.discount,
gae_lambda=args.gae_lambda,
step_size=args.max_kl,
optimizer=ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp(
base_eps=1e-5)) if args.recurrent else None,
mode=args.control,
)
algo.train()
def parse_env_args(self, env, args):
if isinstance(args, dict):
args = to_named_tuple(args)
# Multi-agent wrapper
env = RLLabEnv(env, ma_mode=args.control)
env = MATfEnv(env)
# Policy
if args.recurrent:
if args.feature_net:
feature_network = MLP(
name='feature_net',
input_shape=(env.spec.observation_space.flat_dim +
env.spec.action_space.flat_dim, ),
output_dim=args.feature_output,
hidden_sizes=tuple(args.feature_hidden),
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None)
elif args.conv:
strides = tuple(args.conv_strides)
chans = tuple(args.conv_channels)
filts = tuple(args.conv_filters)
assert len(strides) == len(chans) == len(
filts), "strides, chans and filts not equal"
# only discrete actions supported, should be straightforward to extend to continuous
assert isinstance(
env.spec.action_space,
Discrete), "Only discrete action spaces support conv"
feature_network = ConvNetwork(
name='feature_net',
input_shape=env.spec.observation_space.shape,
output_dim=args.feature_output,
conv_filters=chans,
conv_filter_sizes=filts,
conv_strides=strides,
conv_pads=('VALID', ) * len(chans),
hidden_sizes=tuple(args.feature_hidden),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None)
else:
feature_network = None
if args.recurrent == 'gru':
if isinstance(env.spec.action_space, Box):
if args.control == 'concurrent':
policies = [
GaussianGRUPolicy(env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(
args.policy_hidden[0]),
name='policy_{}'.format(agid))
for agid in range(len(env.agents))
]
policy = GaussianGRUPolicy(env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(
args.policy_hidden[0]),
name='policy')
elif isinstance(env.spec.action_space, Discrete):
if args.control == 'concurrent':
policies = [
CategoricalGRUPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden[0]),
name='policy_{}'.format(agid),
state_include_action=False
if args.conv else True)
for agid in range(len(env.agents))
]
q_network = CategoricalGRUPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden[0]),
name='q_network',
state_include_action=False if args.conv else True)
target_q_network = CategoricalGRUPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden[0]),
name='target_q_network',
state_include_action=False if args.conv else True)
policy = {
'q_network': q_network,
'target_q_network': target_q_network
}
else:
raise NotImplementedError(env.spec.observation_space)
elif args.recurrent == 'lstm':
if isinstance(env.spec.action_space, Box):
if args.control == 'concurrent':
policies = [
GaussianLSTMPolicy(env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(
args.policy_hidden),
name='policy_{}'.format(agid))
for agid in range(len(env.agents))
]
policy = GaussianLSTMPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden),
name='policy')
elif isinstance(env.spec.action_space, Discrete):
if args.control == 'concurrent':
policies = [
CategoricalLSTMPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden),
name='policy_{}'.format(agid))
for agid in range(len(env.agents))
]
q_network = CategoricalLSTMPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden),
name='q_network')
target_q_network = CategoricalLSTMPolicy(
env_spec=env.spec,
feature_network=feature_network,
hidden_dim=int(args.policy_hidden),
name='target_q_network')
policy = {
'q_network': q_network,
'target_q_network': target_q_network
}
else:
raise NotImplementedError(env.spec.action_space)
else:
raise NotImplementedError(args.recurrent)
elif args.conv:
strides = tuple(args.conv_strides)
chans = tuple(args.conv_channels)
filts = tuple(args.conv_filters)
assert len(strides) == len(chans) == len(
filts), "strides, chans and filts not equal"
# only discrete actions supported, should be straightforward to extend to continuous
assert isinstance(
env.spec.action_space,
Discrete), "Only discrete action spaces support conv"
feature_network = ConvNetwork(
name='feature_net',
input_shape=env.spec.observation_space.shape,
output_dim=env.spec.action_space.n,
conv_filters=chans,
conv_filter_sizes=filts,
conv_strides=strides,
conv_pads=(args.conv_pads, ) * len(chans),
hidden_sizes=tuple(args.policy_hidden),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax,
batch_normalization=args.batch_normalization)
if args.algo == 'dqn':
q_network = CategoricalMLPPolicy(name='q_network',
env_spec=env.spec,
prob_network=feature_network)
target_q_network = CategoricalMLPPolicy(
name='target_q_network',
env_spec=env.spec,
prob_network=feature_network)
policy = {
'q_network': q_network,
'target_q_network': target_q_network
}
else:
policy = CategoricalMLPPolicy(name='policy',
env_spec=env.spec,
prob_network=feature_network)
else:
if env.spec is None:
networks = [
DQNNetwork(i,
env,
target_network_update_freq=self.args.
target_network_update,
discount_factor=self.args.discount,
batch_size=self.args.batch_size,
learning_rate=self.args.qfunc_lr)
for i in range(env.n)
]
policy = networks
elif isinstance(env.spec.action_space, Box):
policy = GaussianMLPPolicy(env_spec=env.spec,
hidden_sizes=tuple(
args.policy_hidden),
min_std=args.min_std,
name='policy')
elif isinstance(env.spec.action_space, Discrete):
policy = CategoricalMLPPolicy(env_spec=env.spec,
hidden_sizes=tuple(
args.policy_hidden),
name='policy')
else:
raise NotImplementedError(env.spec.action_space)
return env, policy
def __init__(self,
name,
env_spec,
conv_filters,
conv_filter_sizes,
conv_strides,
conv_pads,
hidden_sizes=(32, 32),
learn_std=True,
init_std=1.0,
adaptive_std=False,
std_share_network=False,
std_hidden_sizes=(32, 32),
min_std=1e-6,
std_hidden_nonlinearity=tf.nn.tanh,
hidden_nonlinearity=tf.nn.tanh,
output_nonlinearity=None,
mean_network=None,
std_network=None,
std_parametrization='exp'):
"""
:param env_spec:
:param hidden_sizes: list of sizes for the fully-connected hidden layers
:param learn_std: Is std trainable
:param init_std: Initial std
:param adaptive_std:
:param std_share_network:
:param std_hidden_sizes: list of sizes for the fully-connected layers for std
:param min_std: whether to make sure that the std is at least some threshold value, to avoid numerical issues
:param std_hidden_nonlinearity:
:param hidden_nonlinearity: nonlinearity used for each hidden layer
:param output_nonlinearity: nonlinearity for the output layer
:param mean_network: custom network for the output mean
:param std_network: custom network for the output log std
:param std_parametrization: how the std should be parametrized. There are a few options:
- exp: the logarithm of the std will be stored, and applied a exponential transformation
- softplus: the std will be computed as log(1+exp(x))
:return:
"""
Serializable.quick_init(self, locals())
assert isinstance(env_spec.action_space, Box)
with tf.variable_scope(name):
obs_dim = env_spec.observation_space.flat_dim
action_dim = env_spec.action_space.flat_dim
# create network
if mean_network is None:
mean_network = ConvNetwork(
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.flat_dim,
conv_filters=conv_filters,
conv_filter_sizes=conv_filter_sizes,
conv_strides=conv_strides,
conv_pads=conv_pads,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=self.linear,
name="prob_network",
)
self._mean_network = mean_network
l_mean = mean_network.output_layer
obs_var = mean_network.input_layer.input_var
if std_network is not None:
l_std_param = std_network.output_layer
else:
if adaptive_std:
std_network = ConvNetwork(
input_shape=env_spec.observation_space.shape,
output_dim=env_spec.action_space.n,
conv_filters=conv_filters,
conv_filter_sizes=conv_filter_sizes,
conv_strides=conv_strides,
conv_pads=conv_pads,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=self.linear,
name="std_network",
)
l_std_param = std_network.output_layer
else:
if std_parametrization == 'exp':
init_std_param = np.log(init_std)
elif std_parametrization == 'softplus':
init_std_param = np.log(np.exp(init_std) - 1)
else:
raise NotImplementedError
l_std_param = L.ParamLayer(
mean_network.input_layer,
num_units=action_dim,
param=tf.constant_initializer(init_std_param),
name="output_std_param",
trainable=learn_std,
)
self.std_parametrization = std_parametrization
if std_parametrization == 'exp':
min_std_param = np.log(min_std)
elif std_parametrization == 'softplus':
min_std_param = np.log(np.exp(min_std) - 1)
else:
raise NotImplementedError
self.min_std_param = min_std_param
# mean_var, log_std_var = L.get_output([l_mean, l_std_param])
#
# if self.min_std_param is not None:
# log_std_var = tf.maximum(log_std_var, np.log(min_std))
#
# self._mean_var, self._log_std_var = mean_var, log_std_var
self._l_mean = l_mean
self._l_std_param = l_std_param
self._dist = DiagonalGaussian(action_dim)
LayersPowered.__init__(self, [l_mean, l_std_param])
super(GaussianConvPolicy, self).__init__(env_spec)
dist_info_sym = self.dist_info_sym(
mean_network.input_layer.input_var, dict())
mean_var = dist_info_sym["mean"]
log_std_var = dist_info_sym["log_std"]
self._f_dist = tensor_utils.compile_function(
inputs=[obs_var],
outputs=[mean_var, log_std_var],
)
def main(_):
env = TfEnv(
AtariEnv(args.env,
force_reset=True,
record_video=False,
record_log=False,
resize_size=args.resize_size,
atari_noop=args.atari_noop,
atari_eplife=args.atari_eplife,
atari_firereset=args.atari_firereset))
policy_network = ConvNetwork(
name='prob_network',
input_shape=env.observation_space.shape,
output_dim=env.action_space.n,
# number of channels/filters for each conv layer
conv_filters=(16, 32),
# filter size
conv_filter_sizes=(8, 4),
conv_strides=(4, 2),
conv_pads=('VALID', 'VALID'),
hidden_sizes=(256, ),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=tf.nn.softmax,
batch_normalization=False)
policy = CategoricalMLPPolicy(name='policy',
env_spec=env.spec,
prob_network=policy_network)
if (args.value_function == 'zero'):
baseline = ZeroBaseline(env.spec)
else:
value_network = get_value_network(env)
baseline_batch_size = args.batch_size * 10
if (args.value_function == 'conj'):
baseline_optimizer = ConjugateGradientOptimizer(
subsample_factor=1.0, num_slices=args.num_slices)
elif (args.value_function == 'adam'):
baseline_optimizer = FirstOrderOptimizer(
max_epochs=3,
batch_size=512,
num_slices=args.num_slices,
verbose=True)
else:
logger.log("Inappropirate value function")
exit(0)
'''
baseline = GaussianMLPBaseline(
env.spec,
num_slices=args.num_slices,
regressor_args=dict(
step_size=0.01,
mean_network=value_network,
optimizer=baseline_optimizer,
subsample_factor=1.0,
batchsize=baseline_batch_size,
use_trust_region=False
)
)
'''
baseline = DeterministicMLPBaseline(env.spec,
num_slices=args.num_slices,
regressor_args=dict(
network=value_network,
optimizer=baseline_optimizer,
normalize_inputs=False))
algo = TRPO(env=env,
policy=policy,
baseline=baseline,
batch_size=args.batch_size,
max_path_length=4500,
n_itr=args.n_itr,
discount=args.discount_factor,
step_size=args.step_size,
clip_reward=(not args.reward_no_scale),
optimizer_args={
"subsample_factor": 1.0,
"num_slices": args.num_slices
}
# plot=True
)
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=args.n_cpu,
inter_op_parallelism_threads=args.n_cpu)
config.gpu_options.allow_growth = True # pylint: disable=E1101
sess = tf.Session(config=config)
sess.__enter__()
algo.train(sess)