def train(self):
self.start_worker()
self.init_opt()
for itr in range(self.current_itr, self.n_itr):
with logger.prefix('itr #%d | ' % itr):
paths = self.sampler.obtain_samples(itr)
samples_data = self.sampler.process_samples(itr, paths)
self.log_diagnostics(paths)
self.optimize_policy(itr, samples_data)
logger.log("saving snapshot...")
params = self.get_itr_snapshot(itr, samples_data)
self.current_itr = itr + 1
params["algo"] = self
if self.store_paths:
params["paths"] = samples_data["paths"]
logger.save_itr_params(itr, params)
logger.log("saved")
logger.dump_tabular(with_prefix=False)
if self.plot:
self.update_plot()
if self.pause_for_plot:
input("Plotting evaluation run: Press Enter to "
"continue...")
self.shutdown_worker()
def train(self):
# This seems like a rather sequential method
pool = SimpleReplayPool(
max_pool_size=self.replay_pool_size,
observation_dim=self.env.observation_space.flat_dim,
action_dim=self.env.action_space.flat_dim,
)
self.start_worker()
self.init_opt()
itr = 0
path_length = 0
path_return = 0
terminal = False
observation = self.env.reset()
sample_policy = pickle.loads(pickle.dumps(self.policy))
for epoch in range(self.n_epochs):
logger.push_prefix('epoch #%d | ' % epoch)
logger.log("Training started")
for epoch_itr in pyprind.prog_bar(range(self.epoch_length)):
# Execute policy
if terminal: # or path_length > self.max_path_length:
# Note that if the last time step ends an episode, the very
# last state and observation will be ignored and not added
# to the replay pool
observation = self.env.reset()
self.es.reset()
sample_policy.reset()
self.es_path_returns.append(path_return)
path_length = 0
path_return = 0
action = self.es.get_action(itr,
observation,
policy=sample_policy) # qf=qf)
next_observation, reward, terminal, _ = self.env.step(action)
path_length += 1
path_return += reward
if not terminal and path_length >= self.max_path_length:
terminal = True
# only include the terminal transition in this case if the flag was set
if self.include_horizon_terminal_transitions:
pool.add_sample(observation, action,
reward * self.scale_reward, terminal)
else:
pool.add_sample(observation, action,
reward * self.scale_reward, terminal)
observation = next_observation
if pool.size >= self.min_pool_size:
for update_itr in range(self.n_updates_per_sample):
# Train policy
batch = pool.random_batch(self.batch_size)
self.do_training(itr, batch)
sample_policy.set_param_values(
self.policy.get_param_values())
itr += 1
logger.log("Training finished")
if pool.size >= self.min_pool_size:
self.evaluate(epoch, pool)
params = self.get_epoch_snapshot(epoch)
logger.save_itr_params(epoch, params)
logger.dump_tabular(with_prefix=False)
logger.pop_prefix()
if self.plot:
self.update_plot()
if self.pause_for_plot:
input("Plotting evaluation run: Press Enter to "
"continue...")
self.env.terminate()
self.policy.terminate()
def train(self):
cur_std = self.sigma0
cur_mean = self.policy.get_param_values()
es = cma_es_lib.CMAEvolutionStrategy(cur_mean, cur_std)
parallel_sampler.populate_task(self.env, self.policy)
if self.plot:
plotter.init_plot(self.env, self.policy)
cur_std = self.sigma0
cur_mean = self.policy.get_param_values()
itr = 0
while itr < self.n_itr and not es.stop():
if self.batch_size is None:
# Sample from multivariate normal distribution.
xs = es.ask()
xs = np.asarray(xs)
# For each sample, do a rollout.
infos = (stateful_pool.singleton_pool.run_map(
sample_return,
[(x, self.max_path_length, self.discount) for x in xs]))
else:
cum_len = 0
infos = []
xss = []
done = False
while not done:
sbs = stateful_pool.singleton_pool.n_parallel * 2
# Sample from multivariate normal distribution.
# You want to ask for sbs samples here.
xs = es.ask(sbs)
xs = np.asarray(xs)
xss.append(xs)
sinfos = stateful_pool.singleton_pool.run_map(
sample_return,
[(x, self.max_path_length, self.discount) for x in xs])
for info in sinfos:
infos.append(info)
cum_len += len(info['returns'])
if cum_len >= self.batch_size:
xs = np.concatenate(xss)
done = True
break
# Evaluate fitness of samples (negative as it is minimization
# problem).
fs = -np.array([info['returns'][0] for info in infos])
# When batching, you could have generated too many samples compared
# to the actual evaluations. So we cut it off in this case.
xs = xs[:len(fs)]
# Update CMA-ES params based on sample fitness.
es.tell(xs, fs)
logger.push_prefix('itr #%d | ' % itr)
logger.record_tabular('Iteration', itr)
logger.record_tabular('CurStdMean', np.mean(cur_std))
undiscounted_returns = np.array(
[info['undiscounted_return'] for info in infos])
logger.record_tabular('AverageReturn',
np.mean(undiscounted_returns))
logger.record_tabular('StdReturn', np.mean(undiscounted_returns))
logger.record_tabular('MaxReturn', np.max(undiscounted_returns))
logger.record_tabular('MinReturn', np.min(undiscounted_returns))
logger.record_tabular('AverageDiscountedReturn', np.mean(fs))
logger.record_tabular(
'AvgTrajLen',
np.mean([len(info['returns']) for info in infos]))
self.env.log_diagnostics(infos)
self.policy.log_diagnostics(infos)
logger.save_itr_params(
itr, dict(
itr=itr,
policy=self.policy,
env=self.env,
))
logger.dump_tabular(with_prefix=False)
if self.plot:
plotter.update_plot(self.policy, self.max_path_length)
logger.pop_prefix()
# Update iteration.
itr += 1
# Set final params.
self.policy.set_param_values(es.result()[0])
parallel_sampler.terminate_task()
def train(self):
with tf.Session() as sess:
if self.load_policy is not None:
import joblib
self.policy = joblib.load(self.load_policy)['policy']
self.init_opt()
# initialize uninitialized vars (I know, it's ugly)
uninit_vars = []
for var in tf.global_variables():
try:
sess.run(var)
except tf.errors.FailedPreconditionError:
uninit_vars.append(var)
sess.run(tf.variables_initializer(uninit_vars))
#sess.run(tf.initialize_all_variables())
self.start_worker()
start_time = time.time()
for itr in range(self.start_itr, self.n_itr):
itr_start_time = time.time()
with logger.prefix('itr #%d | ' % itr):
logger.log("Obtaining samples...")
paths = self.obtain_samples(itr)
logger.log("Processing samples...")
samples_data = self.process_samples(itr, paths)
logger.log("Logging diagnostics...")
self.log_diagnostics(paths)
logger.log("Optimizing policy...")
self.optimize_policy(itr, samples_data)
#new_param_values = self.policy.get_variable_values(self.policy.all_params)
logger.log("Saving snapshot...")
params = self.get_itr_snapshot(itr,
samples_data) # , **kwargs)
if self.store_paths:
params["paths"] = samples_data["paths"]
logger.save_itr_params(itr, params)
logger.log("Saved")
logger.record_tabular('Time', time.time() - start_time)
logger.record_tabular('ItrTime',
time.time() - itr_start_time)
#import pickle
#with open('paths_itr'+str(itr)+'.pkl', 'wb') as f:
# pickle.dump(paths, f)
# debugging
"""
if itr % 1 == 0:
logger.log("Saving visualization of paths")
import matplotlib.pyplot as plt;
for ind in range(5):
plt.clf(); plt.hold(True)
points = paths[ind]['observations']
plt.plot(points[:,0], points[:,1], '-r', linewidth=2)
plt.xlim([-1.0, 1.0])
plt.ylim([-1.0, 1.0])
plt.legend(['path'])
plt.savefig('/home/cfinn/path'+str(ind)+'.png')
"""
# end debugging
logger.dump_tabular(with_prefix=False)
if self.plot:
self.update_plot()
if self.pause_for_plot:
input("Plotting evaluation run: Press Enter to "
"continue...")
self.shutdown_worker()
def train(self):
# TODO - make this a util
flatten_list = lambda l: [item for sublist in l for item in sublist]
with tf.Session() as sess:
# Code for loading a previous policy. Somewhat hacky because needs to be in sess.
if self.load_policy is not None:
import joblib
self.policy = joblib.load(self.load_policy)['policy']
self.init_opt()
# initialize uninitialized vars (only initialize vars that were not loaded)
uninit_vars = []
for var in tf.global_variables():
# note - this is hacky, may be better way to do this in newer TF.
try:
sess.run(var)
except tf.errors.FailedPreconditionError:
uninit_vars.append(var)
sess.run(tf.variables_initializer(uninit_vars))
self.start_worker()
start_time = time.time()
for itr in range(self.start_itr, self.n_itr):
itr_start_time = time.time()
with logger.prefix('itr #%d | ' % itr):
logger.log("Sampling set of tasks/goals for this meta-batch...")
# sample environment configuration
env = self.env
while not ('sample_env_params' in dir(env) or 'sample_goals' in dir(env)):
env = env._wrapped_env
if 'sample_goals' in dir(env):
learner_env_params = env.sample_goals(self.meta_batch_size)
elif 'sample_env_params':
learner_env_params = env.sample_env_params(self.meta_batch_size)
self.policy.switch_to_init_dist() # Switch to pre-update policy
all_samples_data, all_paths = [], []
for step in range(self.num_grad_updates+1):
#if step > 0:
# import pdb; pdb.set_trace() # test param_vals functions.
logger.log('** Step ' + str(step) + ' **')
logger.log("Obtaining samples...")
paths = self.obtain_samples(itr, reset_args=learner_env_params, log_prefix=str(step))
all_paths.append(paths)
logger.log("Processing samples...")
samples_data = {}
for key in paths.keys(): # the keys are the tasks
# don't log because this will spam the consol with every task.
samples_data[key] = self.process_samples(itr, paths[key], log=False)
all_samples_data.append(samples_data)
# for logging purposes
self.process_samples(itr, flatten_list(paths.values()), prefix=str(step), log=True)
logger.log("Logging diagnostics...")
self.log_diagnostics(flatten_list(paths.values()), prefix=str(step))
if step < self.num_grad_updates:
logger.log("Computing policy updates...")
self.policy.compute_updated_dists(samples_data)
logger.log("Optimizing policy...")
# This needs to take all samples_data so that it can construct graph for meta-optimization.
self.optimize_policy(itr, all_samples_data)
logger.log("Saving snapshot...")
params = self.get_itr_snapshot(itr, all_samples_data[-1]) # , **kwargs)
if self.store_paths:
params["paths"] = all_samples_data[-1]["paths"]
logger.save_itr_params(itr, params)
logger.log("Saved")
logger.record_tabular('Time', time.time() - start_time)
logger.record_tabular('ItrTime', time.time() - itr_start_time)
logger.dump_tabular(with_prefix=False)
# The rest is some example plotting code.
# Plotting code is useful for visualizing trajectories across a few different tasks.
if False and itr % 2 == 0 and self.env.observation_space.shape[0] <= 4: # point-mass
logger.log("Saving visualization of paths")
for ind in range(min(5, self.meta_batch_size)):
plt.clf()
plt.plot(learner_env_params[ind][0], learner_env_params[ind][1], 'k*', markersize=10)
plt.hold(True)
preupdate_paths = all_paths[0]
postupdate_paths = all_paths[-1]
pre_points = preupdate_paths[ind][0]['observations']
post_points = postupdate_paths[ind][0]['observations']
plt.plot(pre_points[:,0], pre_points[:,1], '-r', linewidth=2)
plt.plot(post_points[:,0], post_points[:,1], '-b', linewidth=1)
pre_points = preupdate_paths[ind][1]['observations']
post_points = postupdate_paths[ind][1]['observations']
plt.plot(pre_points[:,0], pre_points[:,1], '--r', linewidth=2)
plt.plot(post_points[:,0], post_points[:,1], '--b', linewidth=1)
pre_points = preupdate_paths[ind][2]['observations']
post_points = postupdate_paths[ind][2]['observations']
plt.plot(pre_points[:,0], pre_points[:,1], '-.r', linewidth=2)
plt.plot(post_points[:,0], post_points[:,1], '-.b', linewidth=1)
plt.plot(0,0, 'k.', markersize=5)
plt.xlim([-0.8, 0.8])
plt.ylim([-0.8, 0.8])
plt.legend(['goal', 'preupdate path', 'postupdate path'])
plt.savefig(osp.join(logger.get_snapshot_dir(), 'prepost_path'+str(ind)+'.png'))
elif False and itr % 2 == 0: # swimmer or cheetah
logger.log("Saving visualization of paths")
for ind in range(min(5, self.meta_batch_size)):
plt.clf()
goal_vel = learner_env_params[ind]
plt.title('Swimmer paths, goal vel='+str(goal_vel))
plt.hold(True)
prepathobs = all_paths[0][ind][0]['observations']
postpathobs = all_paths[-1][ind][0]['observations']
plt.plot(prepathobs[:,0], prepathobs[:,1], '-r', linewidth=2)
plt.plot(postpathobs[:,0], postpathobs[:,1], '--b', linewidth=1)
plt.plot(prepathobs[-1,0], prepathobs[-1,1], 'r*', markersize=10)
plt.plot(postpathobs[-1,0], postpathobs[-1,1], 'b*', markersize=10)
plt.xlim([-1.0, 5.0])
plt.ylim([-1.0, 1.0])
plt.legend(['preupdate path', 'postupdate path'], loc=2)
plt.savefig(osp.join(logger.get_snapshot_dir(), 'swim1d_prepost_itr'+str(itr)+'_id'+str(ind)+'.pdf'))
self.shutdown_worker()