def train(self):
plotter = Plotter()
if self.plot:
plotter.init_plot(self.env, self.policy)
self.start_worker()
self.init_opt()
for itr in range(self.current_itr, self.n_itr):
with logger.prefix('itr #{} | '.format(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']
snapshotter.save_itr_params(itr, params)
logger.log('saved')
logger.log(tabular)
if self.plot:
plotter.update_plot(self.policy, self.max_path_length)
if self.pause_for_plot:
input('Plotting evaluation run: Press Enter to '
'continue...')
plotter.close()
self.shutdown_worker()
def train(self,
n_epochs,
n_epoch_cycles=1,
batch_size=None,
plot=False,
store_paths=False,
pause_for_plot=False):
"""Start training.
Args:
n_epochs: Number of epochs.
n_epoch_cycles: Number of batches of samples in each epoch.
This is only useful for off-policy algorithm.
For on-policy algorithm this value should always be 1.
batch_size: Number of steps in batch.
plot: Visualize policy by doing rollout after each epoch.
store_paths: Save paths in snapshot.
pause_for_plot: Pause for plot.
Returns:
The average return in last epoch cycle.
"""
assert self.has_setup, ('Use Runner.setup() to setup runner before '
'training.')
if batch_size is None:
from garage.tf.samplers import OffPolicyVectorizedSampler
if isinstance(self.sampler, OffPolicyVectorizedSampler):
batch_size = self.algo.max_path_length
else:
batch_size = 40 * self.algo.max_path_length
self.n_epoch_cycles = n_epoch_cycles
self.plot = plot
self.start_worker()
self.start_time = time.time()
itr = 0
last_return = None
for epoch in range(n_epochs):
self.itr_start_time = time.time()
paths = None
with logger.prefix('epoch #%d | ' % epoch):
for cycle in range(n_epoch_cycles):
paths = self.obtain_samples(itr, batch_size)
import pdb
pdb.set_trace()
paths = self.sampler.process_samples(itr, paths)
last_return = self.algo.train_once(itr, paths)
itr += 1
self.save_snapshot(epoch, paths if store_paths else None)
self.log_diagnostics(pause_for_plot)
logger.dump_all(itr)
tabular.clear()
self.shutdown_worker()
return last_return
def train(self, sess=None):
address = ("localhost", 6000)
conn = Client(address)
last_average_return = None
try:
created_session = True if (sess is None) else False
if sess is None:
sess = tf.Session()
sess.__enter__()
sess.run(tf.global_variables_initializer())
conn.send(ExpLifecycle.START)
self.start_worker(sess)
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...")
conn.send(ExpLifecycle.OBTAIN_SAMPLES)
paths = self.obtain_samples(itr)
logger.log("Processing samples...")
conn.send(ExpLifecycle.PROCESS_SAMPLES)
samples_data = self.process_samples(itr, paths)
last_average_return = samples_data["average_return"]
logger.log("Logging diagnostics...")
self.log_diagnostics(paths)
logger.log("Optimizing policy...")
conn.send(ExpLifecycle.OPTIMIZE_POLICY)
self.optimize_policy(itr, samples_data)
logger.log("Saving snapshot...")
params = self.get_itr_snapshot(itr)
if self.store_paths:
params["paths"] = samples_data["paths"]
snapshotter.save_itr_params(itr, params)
logger.log("Saved")
tabular.record('Time', time.time() - start_time)
tabular.record('ItrTime', time.time() - itr_start_time)
logger.log(tabular)
if self.plot:
conn.send(ExpLifecycle.UPDATE_PLOT)
self.plotter.update_plot(self.policy,
self.max_path_length)
if self.pause_for_plot:
input("Plotting evaluation run: Press Enter to "
"continue...")
conn.send(ExpLifecycle.SHUTDOWN)
self.shutdown_worker()
if created_session:
sess.close()
finally:
conn.close()
return last_average_return
def train(self):
address = ('localhost', 6000)
conn = Client(address)
try:
plotter = Plotter()
if self.plot:
plotter.init_plot(self.env, self.policy)
conn.send(ExpLifecycle.START)
self.start_worker()
self.init_opt()
for itr in range(self.current_itr, self.n_itr):
with logger.prefix('itr #{} | '.format(itr)):
conn.send(ExpLifecycle.OBTAIN_SAMPLES)
paths = self.sampler.obtain_samples(itr)
conn.send(ExpLifecycle.PROCESS_SAMPLES)
samples_data = self.sampler.process_samples(itr, paths)
self.log_diagnostics(paths)
conn.send(ExpLifecycle.OPTIMIZE_POLICY)
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']
snapshotter.save_itr_params(itr, params)
logger.log('saved')
logger.log(tabular)
if self.plot:
conn.send(ExpLifecycle.UPDATE_PLOT)
plotter.update_plot(self.policy, self.max_path_length)
if self.pause_for_plot:
input('Plotting evaluation run: Press Enter to '
'continue...')
conn.send(ExpLifecycle.SHUTDOWN)
plotter.close()
self.shutdown_worker()
finally:
conn.close()