def optimize_policy(self, all_samples_data, log=True):
"""
Performs MAML outer step
Args:
all_samples_data (list) : list of lists of lists of samples (each is a dict) split by gradient update and
meta task
log (bool) : whether to log statistics
Returns:
None
"""
meta_op_input_dict = self._extract_input_dict_meta_op(all_samples_data, self._optimization_keys)
logger.log("Computing KL before")
mean_kl_before = self.optimizer.constraint_val(meta_op_input_dict)
logger.log("Computing loss before")
loss_before = self.optimizer.loss(meta_op_input_dict)
logger.log("Optimizing")
self.optimizer.optimize(meta_op_input_dict)
logger.log("Computing loss after")
loss_after = self.optimizer.loss(meta_op_input_dict)
logger.log("Computing KL after")
mean_kl = self.optimizer.constraint_val(meta_op_input_dict)
if log:
logger.logkv('MeanKLBefore', mean_kl_before)
logger.logkv('MeanKL', mean_kl)
logger.logkv('LossBefore', loss_before)
logger.logkv('LossAfter', loss_after)
logger.logkv('dLoss', loss_before - loss_after)
def optimize_policy(self, all_samples_data, log=True):
"""
Performs MAML outer step
Args:
all_samples_data (list) : list of lists of lists of samples (each is a dict) split by gradient update and
meta task
log (bool) : whether to log statistics
Returns:
None
"""
meta_op_input_dict = self._extract_input_dict_meta_op(
all_samples_data, self._optimization_keys)
if log:
logger.log("Optimizing")
loss_before = self.optimizer.optimize(
input_val_dict=meta_op_input_dict)
if log:
logger.log("Computing statistics")
loss_after = self.optimizer.loss(input_val_dict=meta_op_input_dict)
if log:
logger.logkv('LossBefore', loss_before)
logger.logkv('LossAfter', loss_after)
def log_diagnostics(self, paths, prefix=''):
progs = [np.mean(path["env_infos"]["reward_forward"]) for path in paths]
ctrl_cost = [-np.mean(path["env_infos"]["reward_ctrl"]) for path in paths]
logger.logkv(prefix+'AverageForwardReturn', np.mean(progs))
logger.logkv(prefix+'MaxForwardReturn', np.max(progs))
logger.logkv(prefix+'MinForwardReturn', np.min(progs))
logger.logkv(prefix+'StdForwardReturn', np.std(progs))
logger.logkv(prefix + 'AverageCtrlCost', np.mean(ctrl_cost))
def log_diagnostics(self, paths, prefix=''):
reach_rew = [path["env_infos"]['reachRew'] for path in paths]
pick_rew = [path["env_infos"]['pickRew'][-1] for path in paths]
place_rew = [path["env_infos"]['placeRew'] for path in paths]
reach_dist = [path["env_infos"]['reachDist'] for path in paths]
placing_dist = [path["env_infos"]['placingDist'] for path in paths]
logger.logkv(prefix + 'AverageReachReward', np.mean(reach_rew))
logger.logkv(prefix + 'AveragePickReward', np.mean(pick_rew))
logger.logkv(prefix + 'AveragePlaceReward', np.mean(place_rew))
logger.logkv(prefix + 'AverageReachDistance', np.mean(reach_dist))
logger.logkv(prefix + 'AveragePlaceDistance', np.mean(placing_dist))
def optimize_policy(self, all_samples_data, log=True):
"""
Performs MAML outer step
Args:
all_samples_data (list) : list of lists of lists of samples (each is a dict) split by gradient update and
meta task
log (bool) : whether to log statistics
Returns:
None
"""
meta_op_input_dict = self._extract_input_dict_meta_op(
all_samples_data, self._optimization_keys)
# add kl_coeffs / clip_eps to meta_op_input_dict
meta_op_input_dict['inner_kl_coeff'] = self.inner_kl_coeff
if self.clip_outer:
meta_op_input_dict['clip_eps'] = self.clip_eps
else:
meta_op_input_dict['outer_kl_coeff'] = self.outer_kl_coeff
if log: logger.log("Optimizing")
loss_before = self.optimizer.optimize(
input_val_dict=meta_op_input_dict)
if log: logger.log("Computing statistics")
loss_after, inner_kls, outer_kl = self.optimizer.compute_stats(
input_val_dict=meta_op_input_dict)
if self.adaptive_inner_kl_penalty:
if log: logger.log("Updating inner KL loss coefficients")
self.inner_kl_coeff = self.adapt_kl_coeff(self.inner_kl_coeff,
inner_kls,
self.target_inner_step)
if self.adaptive_outer_kl_penalty:
if log: logger.log("Updating outer KL loss coefficients")
self.outer_kl_coeff = self.adapt_kl_coeff(self.outer_kl_coeff,
outer_kl,
self.target_outer_step)
if log:
logger.logkv('LossBefore', loss_before)
logger.logkv('LossAfter', loss_after)
logger.logkv('KLInner', np.mean(inner_kls))
logger.logkv('KLCoeffInner', np.mean(self.inner_kl_coeff))
if not self.clip_outer: logger.logkv('KLOuter', outer_kl)
def log_diagnostics(self, paths, prefix=''):
reach_dist = [path["env_infos"]['reachDist'] for path in paths]
placing_dist = [path["env_infos"]['placeDist'] for path in paths]
cos_dist = [path["env_infos"]['cosDist'] for path in paths]
logger.logkv(prefix + 'AverageReachDistance', np.mean(reach_dist))
logger.logkv(prefix + 'AveragePlaceDistance', np.mean(placing_dist))
logger.logkv(prefix + 'AverageCosDistance', np.mean(cos_dist))
def log_diagnostics(self, paths, prefix=''):
fwrd_vel = [path["env_infos"]['reward_run'] for path in paths]
final_fwrd_vel = [path["env_infos"]['reward_run'][-1] for path in paths]
ctrl_cost = [-path["env_infos"]['reward_ctrl'] for path in paths]
logger.logkv(prefix + 'AvgForwardVel', np.mean(fwrd_vel))
logger.logkv(prefix + 'AvgFinalForwardVel', np.mean(final_fwrd_vel))
logger.logkv(prefix + 'AvgCtrlCost', np.std(ctrl_cost))
def _log_path_stats(self, paths, log=False, log_prefix=''):
# compute log stats
average_discounted_return = [sum(path["discounted_rewards"]) for path in paths]
undiscounted_returns = [sum(path["rewards"]) for path in paths]
if log == 'reward':
logger.logkv(log_prefix + 'AverageReturn', np.mean(undiscounted_returns))
elif log == 'all' or log is True:
logger.logkv(log_prefix + 'AverageDiscountedReturn', np.mean(average_discounted_return))
logger.logkv(log_prefix + 'AverageReturn', np.mean(undiscounted_returns))
logger.logkv(log_prefix + 'NumTrajs', len(paths))
logger.logkv(log_prefix + 'StdReturn', np.std(undiscounted_returns))
logger.logkv(log_prefix + 'MaxReturn', np.max(undiscounted_returns))
logger.logkv(log_prefix + 'MinReturn', np.min(undiscounted_returns))
def train(self, tester):
"""
Trains policy on env using algo
Pseudocode:
for itr in n_itr:
for step in num_inner_grad_steps:
sampler.sample()
algo.compute_updated_dists()
algo.optimize_policy()
sampler.update_goals()
"""
best_train_reward = -np.inf
with self.sess.as_default() as sess:
# initialize uninitialized vars (only initialize vars that were not loaded)
uninit_vars = [
var for var in tf.compat.v1.global_variables()
if not sess.run(tf.compat.v1.is_variable_initialized(var))
]
sess.run(tf.compat.v1.variables_initializer(uninit_vars))
start_time = time.time()
for itr in range(self.start_itr, self.n_itr):
itr_start_time = time.time()
logger.log(
"\n ---------------- Iteration %d ----------------" % itr)
logger.log(
"Sampling set of tasks/goals for this meta-batch...")
self.sampler.update_tasks()
self.policy.switch_to_pre_update(
) # Switch to pre-update policy
all_samples_data, all_paths = [], []
list_sampling_time, list_inner_step_time, list_outer_step_time, list_proc_samples_time = [], [], [], []
start_total_inner_time = time.time()
for step in range(self.num_inner_grad_steps + 1):
logger.log('** Step ' + str(step) + ' **')
""" -------------------- Sampling --------------------------"""
logger.log("Obtaining samples...")
time_env_sampling_start = time.time()
paths = self.sampler.obtain_samples(log=True,
log_prefix='Step_%d-' %
step)
list_sampling_time.append(time.time() -
time_env_sampling_start)
all_paths.append(paths)
""" ----------------- Processing Samples ---------------------"""
logger.log("Processing samples...")
time_proc_samples_start = time.time()
samples_data = self.sample_processor.process_samples(
paths, log='all', log_prefix='Step_%d-' % step)
all_samples_data.append(samples_data)
list_proc_samples_time.append(time.time() -
time_proc_samples_start)
self.log_diagnostics(sum(list(paths.values()), []),
prefix='Step_%d-' % step)
""" ------------------- Inner Policy Update --------------------"""
time_inner_step_start = time.time()
if step < self.num_inner_grad_steps:
logger.log("Computing inner policy updates...")
self.algo._adapt(samples_data)
# train_writer = tf.summary.FileWriter('/home/ignasi/Desktop/maml_zoo_graph',
# sess.graph)
list_inner_step_time.append(time.time() -
time_inner_step_start)
total_inner_time = time.time() - start_total_inner_time
time_maml_opt_start = time.time()
""" ------------------ Outer Policy Update ---------------------"""
logger.log("Optimizing policy...")
# This needs to take all samples_data so that it can construct graph for meta-optimization.
time_outer_step_start = time.time()
self.algo.optimize_policy(all_samples_data)
""" ------------------- Logging Stuff --------------------------"""
logger.logkv('Itr', itr)
logger.logkv('n_timesteps',
self.sampler.total_timesteps_sampled)
logger.logkv('Time-OuterStep',
time.time() - time_outer_step_start)
logger.logkv('Time-TotalInner', total_inner_time)
logger.logkv('Time-InnerStep', np.sum(list_inner_step_time))
logger.logkv('Time-SampleProc', np.sum(list_proc_samples_time))
logger.logkv('Time-Sampling', np.sum(list_sampling_time))
logger.logkv('Time', time.time() - start_time)
logger.logkv('ItrTime', time.time() - itr_start_time)
logger.logkv('Time-MAMLSteps',
time.time() - time_maml_opt_start)
logger.log("Saving snapshot...")
params = self.get_itr_snapshot(itr)
logger.save_itr_params(itr, params)
logger.log("Saved")
logger.dumpkvs()
if itr == 0:
sess.graph.finalize()
if (itr % 80 == 79) or (
(self.sample_processor.avg_return >
best_train_reward * 1.05) and
(itr > 50)) and self.sample_processor.avg_return > 300.0:
if self.sample_processor.avg_return > best_train_reward:
best_train_reward = self.sample_processor.avg_return
print('TESTING')
delim = '====================\n'
print(delim, delim, delim)
val_reward = tester.validate(sess)
if self.best_val_reward < val_reward:
self.best_itr = itr
self.best_val_reward = val_reward
print(delim, delim, delim)
logger.log("Training finished")
logger.log(
f"Best iteration is {self.best_itr} with reward {self.best_val_reward}"
)
self.sess.close()
return self.best_itr
def obtain_samples(self, log=False, log_prefix=''):
"""
Collect batch_size trajectories from each task
Args:
log (boolean): whether to log sampling times
log_prefix (str) : prefix for logger
Returns:
(dict) : A dict of paths of size [meta_batch_size] x (batch_size) x [5] x (max_path_length)
"""
# initial setup / preparation
paths = OrderedDict()
for i in range(self.meta_batch_size):
paths[i] = []
n_samples = 0
running_paths = [
_get_empty_running_paths_dict()
for _ in range(self.vec_env.num_envs)
]
pbar = ProgBar(self.total_samples)
policy_time, env_time = 0, 0
policy = self.policy
policy.reset(dones=[True] * self.meta_batch_size)
# initial reset of envs
obses = self.vec_env.reset()
while n_samples < self.total_samples:
# execute policy
t = time.time()
obs_per_task = np.split(np.asarray(obses), self.meta_batch_size)
actions, agent_infos = policy.get_actions(obs_per_task)
policy_time += time.time() - t
# step environments
t = time.time()
actions = np.concatenate(actions) # stack meta batch
next_obses, rewards, dones, env_infos = self.vec_env.step(actions)
env_time += time.time() - t
# stack agent_infos and if no infos were provided (--> None) create empty dicts
agent_infos, env_infos = self._handle_info_dicts(
agent_infos, env_infos)
new_samples = 0
for idx, observation, action, reward, env_info, agent_info, done in zip(
itertools.count(), obses, actions, rewards, env_infos,
agent_infos, dones):
# append new samples to running paths
if isinstance(reward, np.ndarray):
reward = reward[0]
running_paths[idx]["observations"].append(observation)
running_paths[idx]["actions"].append(action)
running_paths[idx]["rewards"].append(reward)
running_paths[idx]["dones"].append(done)
running_paths[idx]["env_infos"].append(env_info)
running_paths[idx]["agent_infos"].append(agent_info)
# if running path is done, add it to paths and empty the running path
if done:
paths[idx // self.envs_per_task].append(
dict(
observations=np.asarray(
running_paths[idx]["observations"]),
actions=np.asarray(running_paths[idx]["actions"]),
rewards=np.asarray(running_paths[idx]["rewards"]),
dones=np.asarray(running_paths[idx]["dones"]),
env_infos=utils.stack_tensor_dict_list(
running_paths[idx]["env_infos"]),
agent_infos=utils.stack_tensor_dict_list(
running_paths[idx]["agent_infos"]),
))
new_samples += len(running_paths[idx]["rewards"])
running_paths[idx] = _get_empty_running_paths_dict()
pbar.update(new_samples)
n_samples += new_samples
obses = next_obses
pbar.stop()
self.total_timesteps_sampled += self.total_samples
if log:
logger.logkv(log_prefix + "PolicyExecTime", policy_time)
logger.logkv(log_prefix + "EnvExecTime", env_time)
return paths
def train(self):
"""
Trains policy on env using algo
Pseudocode:
for itr in n_itr:
for step in num_inner_grad_steps:
sampler.sample()
algo.compute_updated_dists()
algo.optimize_policy()
sampler.update_goals()
"""
with self.sess.as_default() as sess:
# initialize uninitialized vars (only initialize vars that were not loaded)
uninit_vars = [
var for var in tf.compat.v1.global_variables()
if not sess.run(tf.compat.v1.is_variable_initialized(var))
]
sess.run(tf.compat.v1.variables_initializer(uninit_vars))
start_time = time.time()
for itr in range(self.start_itr, self.n_itr):
self.task = self.env.sample_tasks(self.sampler.meta_batch_size)
self.sampler.set_tasks(self.task)
itr_start_time = time.time()
logger.log(
"\n ---------------- Iteration %d ----------------" % itr)
logger.log(
"Sampling set of tasks/goals for this meta-batch...")
""" -------------------- Sampling --------------------------"""
logger.log("Obtaining samples...")
time_env_sampling_start = time.time()
paths = self.sampler.obtain_samples(log=True,
log_prefix='train-')
sampling_time = time.time() - time_env_sampling_start
""" ----------------- Processing Samples ---------------------"""
logger.log("Processing samples...")
time_proc_samples_start = time.time()
samples_data = self.sample_processor.process_samples(
paths, log='all', log_prefix='train-')
proc_samples_time = time.time() - time_proc_samples_start
self.log_diagnostics(sum(paths.values(), []), prefix='train-')
""" ------------------ Policy Update ---------------------"""
logger.log("Optimizing policy...")
# This needs to take all samples_data so that it can construct graph for meta-optimization.
time_optimization_step_start = time.time()
self.algo.optimize_policy(samples_data)
""" ------------------- Logging Stuff --------------------------"""
logger.logkv('Itr', itr)
logger.logkv('n_timesteps',
self.sampler.total_timesteps_sampled)
logger.logkv('Time-Optimization',
time.time() - time_optimization_step_start)
logger.logkv('Time-SampleProc', np.sum(proc_samples_time))
logger.logkv('Time-Sampling', sampling_time)
logger.logkv('Time', time.time() - start_time)
logger.logkv('ItrTime', time.time() - itr_start_time)
logger.log("Saving snapshot...")
params = self.get_itr_snapshot(itr)
logger.save_itr_params(itr, params)
logger.log("Saved")
logger.dumpkvs()
if itr == 0:
sess.graph.finalize()
logger.log("Training finished")
self.sess.close()
def log_diagnostics(self, paths, prefix=''):
for key in self.info_logKeys:
logger.logkv(prefix + 'last_'+key, np.mean([path['env_infos'][key][-1] for path in paths]) )
def log_diagnostics(self, paths, prefix=''):
"""
Log extra information per iteration based on the collected paths
"""
log_stds = np.vstack([path["agent_infos"]["log_std"] for path in paths])
logger.logkv(prefix+'AveragePolicyStd', np.mean(np.exp(log_stds)))