def __init__(
self,
obs_spec: Spec,
act_spec: Spec,
model_fn: ModelBuilder = None,
policy_cls: PolicyType = None,
sess_mgr: SessionManager = None,
optimizer: tf.train.Optimizer = None,
value_coef=DEFAULTS['value_coef'],
entropy_coef=DEFAULTS['entropy_coef'],
traj_len=DEFAULTS['traj_len'],
batch_sz=DEFAULTS['batch_sz'],
discount=DEFAULTS['discount'],
gae_lambda=DEFAULTS['gae_lambda'],
clip_rewards=DEFAULTS['clip_rewards'],
clip_grads_norm=DEFAULTS['clip_grads_norm'],
normalize_returns=DEFAULTS['normalize_returns'],
normalize_advantages=DEFAULTS['normalize_advantages'],
):
MemoryAgent.__init__(self, obs_spec, act_spec, traj_len, batch_sz)
if not sess_mgr:
sess_mgr = SessionManager()
if not optimizer:
optimizer = tf.train.AdamOptimizer(
learning_rate=DEFAULTS['learning_rate'])
self.sess_mgr = sess_mgr
self.value_coef = value_coef
self.entropy_coef = entropy_coef
self.discount = discount
self.gae_lambda = gae_lambda
self.clip_rewards = clip_rewards
self.normalize_returns = normalize_returns
self.normalize_advantages = normalize_advantages
self.model = model_fn(obs_spec, act_spec)
self.value = self.model.outputs[-1]
self.policy = policy_cls(act_spec, self.model.outputs[:-1])
self.loss_op, self.loss_terms, self.loss_inputs = self.loss_fn()
# noinspection PyShadowingBuiltins
grads, vars = zip(*optimizer.compute_gradients(self.loss_op))
self.grads_norm = tf.global_norm(grads)
if clip_grads_norm > 0.:
grads, _ = tf.clip_by_global_norm(grads, clip_grads_norm,
self.grads_norm)
self.train_op = optimizer.apply_gradients(
zip(grads, vars), global_step=sess_mgr.global_step)
self.minimize_ops = self.make_minimize_ops()
sess_mgr.restore_or_init()
self.n_batches = sess_mgr.start_step
self.start_step = sess_mgr.start_step * traj_len
self.logger = Logger()
def __init__(self,
env_spec,
callbacks=None,
model_class=FullyConvModel,
optimizer=tf.train.AdamOptimizer,
learning_rate=0.0001,
discount=0.99,
trajectory_length=16,
batch_size=32,
max_grads_norm=100,
policy_factor=1,
entropy_factor=0.0001,
value_factor=0.5):
self.callbacks = callbacks
self.discount = discount
self.policy_factor = policy_factor
self.entropy_factor = entropy_factor
self.value_factor = value_factor
self.input_observations = {
name: Input(shape=spec.shape, name='input_{}'.format(name))
for name, spec in env_spec.observation_spec.items()
}
self.input_actions = {
name: Input(shape=(),
name='input_arg_{}_value'.format(name),
dtype='int32')
for name in env_spec.action_spec
}
self.input_returns = Input(shape=(), name='input_returns')
self.function_args_mask = tf.constant(
env_spec.action_spec['function_id'].args_mask,
dtype=tf.float32,
name='function_args_mask')
self.model = model_class(self.input_observations, env_spec)
self.loss = self.build_loss()
self.optimizer = optimizer(learning_rate=learning_rate)
grads, vars = zip(*self.optimizer.compute_gradients(self.loss))
grads_norm = tf.global_norm(grads)
if max_grads_norm > 0:
grads, _ = tf.clip_by_global_norm(grads, max_grads_norm,
grads_norm)
self.train_op = self.optimizer.apply_gradients(
zip(grads, vars), global_step=tf.train.get_or_create_global_step())
self.history = History(trajectory_length, batch_size, env_spec)
tf.summary.scalar('learning_rate', learning_rate)
tf.summary.scalar('total_loss', self.loss, family='losses')
tf.summary.scalar('grads_norm', grads_norm)
def __init__(
self,
obs_spec: Spec, # how does this just work like that?
act_spec: Spec,
model_fn: ModelBuilder=None, # same with these. Has to do with gin and/or __init__ files
policy_cls: PolicyType=None,
sess_mgr: SessionManager=None,
optimizer: tf.train.Optimizer=None,
value_coef=DEFAULTS['value_coef'],
entropy_coef=DEFAULTS['entropy_coef'],
traj_len=DEFAULTS['traj_len'],
batch_sz=DEFAULTS['batch_sz'],
gamma=DEFAULTS['gamma'],
gae_lambda=DEFAULTS['gae_lambda'],
clip_rewards=DEFAULTS['clip_rewards'],
clip_grads_norm=DEFAULTS['clip_grads_norm'],
normalize_returns=DEFAULTS['normalize_returns'],
normalize_advantages=DEFAULTS['normalize_advantages']):
MemoryAgent.__init__(self, obs_spec, act_spec, traj_len, batch_sz)
if not sess_mgr:
sess_mgr = SessionManager()
if not optimizer:
optimizer = tf.train.AdamOptimizer(DEFAULTS["learning_rate"])
self.sess_mgr = sess_mgr
self.value_coef = value_coef
self.entropy_coef = entropy_coef
self.traj_len = traj_len
self.gamma = gamma
self.gae_lambda = gae_lambda
self.clip_rewards = clip_rewards
self.normalize_returns = normalize_returns
self.normalize_advantages = normalize_advantages
self.model = model_fn(obs_sec, act_spec) # this is fully_conv
self.value = self.model.outputs[-1] # very cool
self.policy = policy_cls(act_spec, self.model.outputs[:-1]) # whats advantage of doing it like this? Also Policy(act_spec, logits)
self.loss_op, self.loss_terms, self.loss_inputs = self.loss_fn() # does this have to change?
grads, vars = zip(*optimizer.compute_gradients(self.loss_op))
self.grads_norm = tf.global_norm(grads)
if clip_grads_norm > 0.: # currently defaults at 0.
grads, _ = tf.clip_by_global_norm(grads, clip_grads_norm, self.grads_norm)
self.train_op = optimizer.apply_gradients(zip(grads, vars), global_step=sess_mgr.global_step)
self.minimize_ops = self.make_minimize_ops() # whats benefit of doing like this instead of just setting them in the method?
sess_mgr.restore_or_init()
self.n_batches = sess_mgr.start_step
self.start_step = sess_mgr.start_step * traj_len # why is there no self on these?
self.logger = Logger() # whats this look like exactly?
def __init__(
self,
obs_spec: Spec,
act_spec: Spec,
model_fn: ModelBuilder,
policy_cls: PolicyType,
sess_mgr: SessionManager = None,
traj_len=16,
batch_sz=16,
discount=0.99,
gae_lambda=0.95,
clip_rewards=0.0,
normalize_advantages=True,
bootstrap_terminals=False,
clip_grads_norm=0.0,
optimizer=tf.train.AdamOptimizer(),
logger=Logger()
):
MemoryAgent.__init__(self, obs_spec, act_spec, traj_len, batch_sz)
if not sess_mgr:
sess_mgr = SessionManager()
self.sess_mgr = sess_mgr
self.discount = discount
self.gae_lambda = gae_lambda
self.clip_rewards = clip_rewards
self.normalize_advantages = normalize_advantages
self.bootstrap_terminals = bootstrap_terminals
self.logger = logger
self.model = model_fn(obs_spec, act_spec)
self.value = self.model.outputs[-1]
self.policy = policy_cls(act_spec, self.model.outputs[:-1])
self.loss_op, self.loss_terms, self.loss_inputs = self.loss_fn()
grads, vars = zip(*optimizer.compute_gradients(self.loss_op))
self.grads_norm = tf.global_norm(grads)
if clip_grads_norm > 0.:
grads, _ = tf.clip_by_global_norm(grads, clip_grads_norm, self.grads_norm)
self.train_op = optimizer.apply_gradients(zip(grads, vars), global_step=self.sess_mgr.global_step)
self.sess_mgr.restore_or_init()
# NB! changing trajectory length in-between checkpoints will break the logs
self.n_batches = self.sess_mgr.start_step
self.start_step = self.sess_mgr.start_step * traj_len
def __init__(
self,
obs_spec,
act_spec,
model_fn=build_mlp,
policy_cls=MultiPolicy,
sess_mgr=None,
traj_len=16,
batch_sz=16,
discount=0.99,
gae_lambda=0.95,
clip_rewards=0.0,
normalize_advantages=True,
bootstrap_terminals=False,
clip_grads_norm=0.0,
optimizer=tf.train.AdamOptimizer(),
logger=Logger()
):
MemoryAgent.__init__(self, obs_spec, act_spec, traj_len, batch_sz)
if not sess_mgr:
sess_mgr = SessionManager()
self.sess_mgr = sess_mgr
self.discount = discount
self.gae_lambda = gae_lambda
self.clip_rewards = clip_rewards
self.normalize_advantages = normalize_advantages
self.bootstrap_terminals = bootstrap_terminals
self.logger = logger
self.model = model_fn(obs_spec, act_spec)
self.value = self.model.outputs[-1]
self.policy = policy_cls(act_spec, self.model.outputs[:-1])
self.loss_op, self.loss_terms, self.loss_inputs = self.loss_fn()
grads, vars = zip(*optimizer.compute_gradients(self.loss_op))
self.grads_norm = tf.global_norm(grads)
if clip_grads_norm > 0.:
grads, _ = tf.clip_by_global_norm(grads, clip_grads_norm, self.grads_norm)
self.train_op = optimizer.apply_gradients(zip(grads, vars), global_step=self.sess_mgr.global_step)
self.sess_mgr.restore_or_init()
# NB! changing trajectory length in-between checkpoints will break the logs
self.n_batches = self.sess_mgr.start_step
self.start_step = self.sess_mgr.start_step * traj_len
def __init__(
self,
obs_spec: Spec,
act_spec: Spec,
model_variable_scope=DEFAULTS['model_variable_scope'],
model_fn: ModelBuilder = None,
policy_cls: PolicyType = None,
sess_mgr: SessionManager = None,
optimizer: tf.train.Optimizer = None,
value_coef=DEFAULTS['value_coef'],
entropy_coef=DEFAULTS['entropy_coef'],
traj_len=DEFAULTS['traj_len'],
batch_sz=DEFAULTS['batch_sz'],
discount=DEFAULTS['discount'],
gae_lambda=DEFAULTS['gae_lambda'],
clip_rewards=DEFAULTS['clip_rewards'],
clip_grads_norm=DEFAULTS['clip_grads_norm'],
normalize_returns=DEFAULTS['normalize_returns'],
normalize_advantages=DEFAULTS['normalize_advantages'],
**kwargs,
):
MemoryAgent.__init__(self, obs_spec, act_spec, traj_len, batch_sz)
print(LOGGING_MSG_HEADER +
": the traj_len is {} and batch_sz is {}".format(
traj_len, batch_sz))
if not sess_mgr:
sess_mgr = SessionManager()
self.subenvs = subenvs = kwargs[
'subenvs'] if 'subenvs' in kwargs else []
if optimizer:
optimizers = [copy.deepcopy(optimizer) for subenv in subenvs]
else:
optimizer = tf.train.AdamOptimizer(
learning_rate=DEFAULTS['learning_rate'])
optimizers = [
tf.train.AdamOptimizer(learning_rate=DEFAULTS['learning_rate'])
for subenv in subenvs
]
self.sess_mgr = sess_mgr
self.model_variable_scope = self.sess_mgr.model_variable_scope
self.value_coef = value_coef
self.entropy_coef = entropy_coef
self.discount = discount
self.gae_lambda = gae_lambda
self.clip_rewards = clip_rewards
self.normalize_returns = normalize_returns
self.normalize_advantages = normalize_advantages
self.traj_len = traj_len
self.batch_sz = batch_sz
print(LOGGING_MSG_HEADER + " : the current model_variable_scope is",
self.model_variable_scope)
# implement the a2c to support multiple subagents
# self.model = model_fn(obs_spec, act_spec)
with sess_mgr.sess.graph.as_default():
# note this is name_scope as opposed to variable_scope, important
with tf.name_scope(self.sess_mgr.main_tf_vs.original_name_scope):
if subenvs:
from collections import defaultdict
self.subenv_dict = defaultdict(list)
print(
LOGGING_MSG_HEADER +
": Creating models for each individual subenvs: ",
subenvs)
for i, subenv in enumerate(subenvs):
subenv_model = model_fn(obs_spec, act_spec)
self.subenv_dict['models'].append(subenv_model)
subenv_value = subenv_model.outputs[-1]
self.subenv_dict['values'].append(subenv_value)
subenv_policy = policy_cls(act_spec,
subenv_model.outputs[:-1])
self.subenv_dict['policies'].append(subenv_policy)
subenv_loss_op, subenv_loss_terms, subenv_loss_inputs = self.loss_fn(
policy=subenv_policy, value=subenv_value)
self.subenv_dict['loss_ops'].append(subenv_loss_op)
self.subenv_dict['loss_terms'].append(
subenv_loss_terms)
self.subenv_dict['loss_inputs'].append(
subenv_loss_inputs)
subenv_optimizer = optimizers[i]
grads, vars = zip(*subenv_optimizer.compute_gradients(
subenv_loss_op))
subenv_grads_norm = tf.global_norm(grads)
self.subenv_dict['grads_norms'].append(
subenv_grads_norm)
if clip_grads_norm > 0:
grads, _ = tf.clip_by_global_norm(
grads, clip_grads_norm, subenv_grads_norm)
self.subenv_dict['train_ops'].append(
subenv_optimizer.apply_gradients(
zip(grads, vars),
global_step=sess_mgr.global_step))
self.subenv_dict['minimize_ops'].append(
self.make_minimize_ops(subenv_id=i))
print(
LOGGING_MSG_HEADER +
": Successfully created models for each individual subenvs"
)
else:
print(LOGGING_MSG_HEADER +
": Creating single model for the environment.")
self.model = model_fn(obs_spec, act_spec)
self.value = self.model.outputs[-1]
self.policy = policy_cls(act_spec, self.model.outputs[:-1])
self.loss_op, self.loss_terms, self.loss_inputs = self.loss_fn(
)
grads, vars = zip(
*optimizer.compute_gradients(self.loss_op))
self.grads_norm = tf.global_norm(grads)
if clip_grads_norm > 0.:
grads, _ = tf.clip_by_global_norm(
grads, clip_grads_norm, self.grads_norm)
self.train_op = optimizer.apply_gradients(
zip(grads, vars), global_step=sess_mgr.global_step)
self.minimize_ops = self.make_minimize_ops()
print(LOGGING_MSG_HEADER +
" : main_model setup on sess and graph complete")
sess_mgr.restore_or_init()
print(LOGGING_MSG_HEADER +
" : main_model weights restore/init complete")
self.n_batches = sess_mgr.start_step
self.start_step = sess_mgr.start_step * traj_len
self.logger = Logger()