def __init__(self,
name,
input_dim,
output_dim,
hidden_sizes=(32, 32),
hidden_nonlinearity=tf.nn.relu,
output_nonlinearity=None,
input_var=None,
**kwargs):
Serializable.quick_init(self, locals())
self.input_dim = input_dim
self.output_dim = output_dim
self.name = name
self.input_var = input_var
self.hidden_sizes = hidden_sizes
self.hidden_nonlinearity = hidden_nonlinearity
self.output_nonlinearity = output_nonlinearity
self.batch_normalization = kwargs.get('batch_normalization', False)
self._params = None
self._assign_ops = None
self._assign_phs = None
def __init__(self, *args, **kwargs):
# store the init args for serialization and call the super constructors
Serializable.quick_init(self, locals())
Layer.__init__(self, *args, **kwargs)
self._cell_type = kwargs.get('cell_type', 'gru')
self.state_var = kwargs.get('state_var', None)
self.build_graph()
def __init__(self, *args, **kwargs):
# store the init args for serialization and call the super constructors
Serializable.quick_init(self, locals())
Layer.__init__(self, *args, **kwargs)
self.num_filters = kwargs.get('num_filters')
self.kernel_sizes = kwargs.get('kernel_sizes')
self.strides = kwargs.get('strides')
self.hidden_nonlinearity = kwargs.get('hidden_nonlinearity')
self.output_nonlinearity = kwargs.get('output_nonlinearity')
self.build_graph()
def __init__(
self,
env,
policy,
num_rollouts,
max_path_length,
n_parallel=1,
vae=None,
):
Serializable.quick_init(self, locals())
super(Sampler, self).__init__(env, policy, n_parallel, max_path_length)
self.total_samples = num_rollouts * max_path_length
self.n_parallel = n_parallel
self.total_timesteps_sampled = 0
self.vae = vae
# setup vectorized environment
if self.n_parallel > 1:
self.vec_env = ParallelEnvExecutor(env, n_parallel, num_rollouts, self.max_path_length)
else:
self.vec_env = IterativeEnvExecutor(env, num_rollouts, self.max_path_length)
def __init__(
self,
env,
scale_reward=1.,
normalize_obs=False,
normalize_reward=False,
obs_alpha=0.001,
reward_alpha=0.001,
normalization_scale=1.,
):
Serializable.quick_init(self, locals())
self._scale_reward = 1
self._wrapped_env = env
self._normalize_obs = normalize_obs
self._normalize_reward = normalize_reward
self._obs_alpha = obs_alpha
self._obs_mean = np.zeros(self.observation_space.shape)
self._obs_var = np.ones(self.observation_space.shape)
self._reward_alpha = reward_alpha
self._reward_mean = 0.
self._reward_var = 1.
self._normalization_scale = normalization_scale
def __init__(self, *args, **kwargs):
# store the init args for serialization and call the super constructors
Serializable.quick_init(self, locals())
Layer.__init__(self, *args, **kwargs)
self.build_graph()
def __init__(
self,
name,
env,
hidden_sizes=(500, 500),
hidden_nonlinearity="tanh",
output_nonlinearity=None,
batch_size=500,
learning_rate=0.001,
weight_normalization=True,
normalize_input=True,
optimizer=tf.train.AdamOptimizer,
valid_split_ratio=0.2,
rolling_average_persitency=0.99,
buffer_size=100000,
):
Serializable.quick_init(self, locals())
self.normalization = None
self.normalize_input = normalize_input
self.use_reward_model = False
self.buffer_size = buffer_size
self.name = name
self.hidden_sizes = hidden_sizes
self._dataset_train = None
self._dataset_test = None
self.next_batch = None
self.valid_split_ratio = valid_split_ratio
self.rolling_average_persitency = rolling_average_persitency
self.hidden_nonlinearity = hidden_nonlinearity = self._activations[
hidden_nonlinearity]
self.output_nonlinearity = output_nonlinearity = self._activations[
output_nonlinearity]
with tf.variable_scope(name):
self.batch_size = batch_size
self.learning_rate = learning_rate
# determine dimensionality of state and action space
self.obs_space_dims = env.observation_space.shape[0]
self.action_space_dims = env.action_space.shape[0]
# placeholders
self.obs_ph = tf.placeholder(tf.float32,
shape=(None, self.obs_space_dims))
self.act_ph = tf.placeholder(tf.float32,
shape=(None, self.action_space_dims))
self.delta_ph = tf.placeholder(tf.float32,
shape=(None, self.obs_space_dims))
self._create_stats_vars()
# concatenate action and observation --> NN input
self.nn_input = tf.concat([self.obs_ph, self.act_ph], axis=1)
# create MLP
mlp = MLP(name,
output_dim=self.obs_space_dims,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
input_var=self.nn_input,
input_dim=self.obs_space_dims + self.action_space_dims,
weight_normalization=weight_normalization)
self.delta_pred = mlp.output_var
# define loss and train_op
self.loss = tf.reduce_mean(
tf.linalg.norm(self.delta_ph - self.delta_pred, axis=-1))
self.optimizer = optimizer(self.learning_rate)
self.train_op = self.optimizer.minimize(self.loss)
# tensor_utils
self.f_delta_pred = compile_function([self.obs_ph, self.act_ph],
self.delta_pred)
self._networks = [mlp]
def __init__(
self,
name,
env,
num_models=5,
hidden_sizes=(512, 512),
hidden_nonlinearity='swish',
output_nonlinearity=None,
batch_size=500,
learning_rate=0.001,
weight_normalization=False, # Doesn't work
normalize_input=True,
optimizer=tf.train.AdamOptimizer,
valid_split_ratio=0.2, # 0.1
rolling_average_persitency=0.99,
buffer_size=50000,
loss_str='MSE',
):
Serializable.quick_init(self, locals())
max_logvar = 1
min_logvar = 0.1
self.normalization = None
self.normalize_input = normalize_input
self.next_batch = None
self.valid_split_ratio = valid_split_ratio
self.rolling_average_persitency = rolling_average_persitency
self.buffer_size_train = int(buffer_size * (1 - valid_split_ratio))
self.buffer_size_test = int(buffer_size * valid_split_ratio)
self.batch_size = batch_size
self.learning_rate = learning_rate
self.num_models = num_models
self.hidden_sizes = hidden_sizes
self.name = name
self._dataset_train = None
self._dataset_test = None
# determine dimensionality of state and action space
self.obs_space_dims = obs_space_dims = env.observation_space.shape[0]
self.action_space_dims = action_space_dims = env.action_space.shape[0]
self.timesteps_counter = 0
self.used_timesteps_counter = 0
self.hidden_nonlinearity = hidden_nonlinearity = self._activations[
hidden_nonlinearity]
self.output_nonlinearity = output_nonlinearity = self._activations[
output_nonlinearity]
""" computation graph for training and simple inference """
with tf.variable_scope(name, reuse=tf.AUTO_REUSE):
# placeholders
self.obs_ph = tf.placeholder(tf.float32,
shape=(None, obs_space_dims))
self.act_ph = tf.placeholder(tf.float32,
shape=(None, action_space_dims))
self.delta_ph = tf.placeholder(tf.float32,
shape=(None, obs_space_dims))
self._create_stats_vars()
# concatenate action and observation --> NN input
self.nn_input = tf.concat([self.obs_ph, self.act_ph], axis=1)
obs_ph = tf.split(self.nn_input, self.num_models, axis=0)
# create MLP
mlps = []
delta_preds = []
self.obs_next_pred = []
for i in range(num_models):
with tf.variable_scope('model_{}'.format(i),
reuse=tf.AUTO_REUSE):
mlp = MLP(
name + '/model_{}'.format(i),
output_dim=obs_space_dims,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
input_var=obs_ph[i],
input_dim=obs_space_dims + action_space_dims,
)
mlps.append(mlp)
delta_preds.append(mlp.output_var)
self.delta_pred = tf.stack(
delta_preds, axis=2) # shape: (batch_size, ndim_obs, n_models)
# define loss and train_op
if loss_str == 'L2':
self.loss = tf.reduce_mean(
tf.linalg.norm(self.delta_ph[:, :, None] - self.delta_pred,
axis=1))
elif loss_str == 'MSE':
self.loss = tf.reduce_mean(
(self.delta_ph[:, :, None] - self.delta_pred)**2)
else:
raise NotImplementedError
self.optimizer = optimizer(learning_rate=self.learning_rate)
self.train_op = self.optimizer.minimize(self.loss)
# tensor_utils
self.f_delta_pred = compile_function([self.obs_ph, self.act_ph],
self.delta_pred)
""" computation graph for inference where each of the models receives a different batch"""
with tf.variable_scope(name, reuse=True):
# placeholders
self.obs_model_batches_stack_ph = tf.placeholder(
tf.float32, shape=(None, obs_space_dims))
self.act_model_batches_stack_ph = tf.placeholder(
tf.float32, shape=(None, action_space_dims))
self.delta_model_batches_stack_ph = tf.placeholder(
tf.float32, shape=(None, obs_space_dims))
# split stack into the batches for each model --> assume each model receives a batch of the same size
self.obs_model_batches = tf.split(self.obs_model_batches_stack_ph,
self.num_models,
axis=0)
self.act_model_batches = tf.split(self.act_model_batches_stack_ph,
self.num_models,
axis=0)
self.delta_model_batches = tf.split(
self.delta_model_batches_stack_ph, self.num_models, axis=0)
# reuse previously created MLP but each model receives its own batch
delta_preds = []
self.obs_next_pred = []
self.loss_model_batches = []
self.train_op_model_batches = []
for i in range(num_models):
with tf.variable_scope('model_{}'.format(i), reuse=True):
# concatenate action and observation --> NN input
nn_input = tf.concat(
[self.obs_model_batches[i], self.act_model_batches[i]],
axis=1)
mlp = MLP(name + '/model_{}'.format(i),
output_dim=obs_space_dims,
hidden_sizes=hidden_sizes,
hidden_nonlinearity=hidden_nonlinearity,
output_nonlinearity=output_nonlinearity,
input_var=nn_input,
input_dim=obs_space_dims + action_space_dims,
weight_normalization=weight_normalization)
delta_preds.append(mlp.output_var)
# define loss and train_op
if loss_str == 'L2':
loss = tf.reduce_mean(
tf.linalg.norm(self.delta_model_batches[i] -
mlp.output_var,
axis=1))
elif loss_str == 'MSE':
loss = tf.reduce_mean(
(self.delta_model_batches[i] - mlp.output_var)**2)
else:
raise NotImplementedError
self.loss_model_batches.append(loss)
self.train_op_model_batches.append(
optimizer(learning_rate=self.learning_rate).minimize(loss))
self.delta_pred_model_batches_stack = tf.concat(
delta_preds,
axis=0) # shape: (batch_size_per_model*num_models, ndim_obs)
# tensor_utils
self.f_delta_pred_model_batches = compile_function([
self.obs_model_batches_stack_ph,
self.act_model_batches_stack_ph
], self.delta_pred_model_batches_stack)
self._networks = mlps