How to use the parl.utils.window_stat.WindowStat function in parl
To help you get started, we’ve selected a few parl examples, based on popular ways it is used in public projects.
Secure your code as it's written. Use Snyk Code to scan source code in minutes - no build needed - and fix issues immediately.
PaddlePaddle / PARL / examples / IMPALA / train.py View on Github 
assert get_gpu_count() == 1, 'Only support training in single GPU,\
Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_TO_USE]` .'
self.cache_params = self.agent.get_weights()
self.params_lock = threading.Lock()
self.params_updated = False
self.cache_params_sent_cnt = 0
self.total_params_sync = 0
#========== Learner ==========
self.lr, self.entropy_coeff = None, None
self.lr_scheduler = PiecewiseScheduler(config['lr_scheduler'])
self.entropy_coeff_scheduler = PiecewiseScheduler(
config['entropy_coeff_scheduler'])
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.kl_stat = WindowStat(100)
self.learn_time_stat = TimeStat(100)
self.start_time = None
self.learn_thread = threading.Thread(target=self.run_learn)
self.learn_thread.setDaemon(True)
self.learn_thread.start()
#========== Remote Actor ===========
self.remote_count = 0
self.batch_buffer = []
self.remote_metrics_queue = queue.Queue()if machine_info.is_gpu_available():
assert get_gpu_count() == 1, 'Only support training in single GPU,\
Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_TO_USE]` .'
else:
cpu_num = os.environ.get('CPU_NUM')
assert cpu_num is not None and cpu_num == '1', 'Only support training in single CPU,\
Please set environment variable: `export CPU_NUM=1`.'
#========== Learner ==========
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.lr = None
self.entropy_coeff = None
self.learn_time_stat = TimeStat(100)
self.start_time = None
#========== Remote Actor ===========
self.remote_count = 0
self.sample_data_queue = queue.Queue()
self.remote_metrics_queue = queue.Queue()
self.sample_total_steps = 0
self.params_queues = []
self.create_actors()model, vf_loss_coeff=config['vf_loss_coeff'])
self.agent = AtariAgent(algorithm, config)
if machine_info.is_gpu_available():
assert get_gpu_count() == 1, 'Only support training in single GPU,\
Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_TO_USE]` .'
else:
cpu_num = os.environ.get('CPU_NUM')
assert cpu_num is not None and cpu_num == '1', 'Only support training in single CPU,\
Please set environment variable: `export CPU_NUM=1`.'
#========== Learner ==========
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.lr = None
self.entropy_coeff = None
self.learn_time_stat = TimeStat(100)
self.start_time = None
#========== Remote Actor ===========
self.remote_count = 0
self.sample_data_queue = queue.Queue()
self.remote_metrics_queue = queue.Queue()
self.sample_total_steps = 0
self.params_queues = []self.cache_params = self.agent.get_weights()
self.params_lock = threading.Lock()
self.params_updated = False
self.cache_params_sent_cnt = 0
self.total_params_sync = 0
#========== Learner ==========
self.lr, self.entropy_coeff = None, None
self.lr_scheduler = PiecewiseScheduler(config['lr_scheduler'])
self.entropy_coeff_scheduler = PiecewiseScheduler(
config['entropy_coeff_scheduler'])
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.kl_stat = WindowStat(100)
self.learn_time_stat = TimeStat(100)
self.start_time = None
self.learn_thread = threading.Thread(target=self.run_learn)
self.learn_thread.setDaemon(True)
self.learn_thread.start()
#========== Remote Actor ===========
self.remote_count = 0
self.batch_buffer = []
self.remote_metrics_queue = queue.Queue()
self.sample_total_steps = 0
self.create_actors()assert get_gpu_count() == 1, 'Only support training in single GPU,\
Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_YOU_WANT_TO_USE]` .'
else:
cpu_num = os.environ.get('CPU_NUM')
assert cpu_num is not None and cpu_num == '1', 'Only support training in single CPU,\
Please set environment variable: `export CPU_NUM=1`.'
#========== Learner ==========
self.lr, self.entropy_coeff = None, None
self.lr_scheduler = PiecewiseScheduler(config['lr_scheduler'])
self.entropy_coeff_scheduler = PiecewiseScheduler(
config['entropy_coeff_scheduler'])
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.learn_time_stat = TimeStat(100)
self.start_time = None
# learn thread
self.learn_thread = threading.Thread(target=self.run_learn)
self.learn_thread.setDaemon(True)
self.learn_thread.start()
self.predict_input_queue = queue.Queue()
# predict thread
self.predict_threads = []
for i in six.moves.range(self.config['predict_thread_num']):if args.restore_model_path is not None:
self.restore(args.restore_model_path)
# add lock between training and predicting
self.model_lock = threading.Lock()
# add lock when appending data to rpm or writing scalars to tensorboard
self.memory_lock = threading.Lock()
self.ready_actor_queue = queue.Queue()
self.total_steps = 0
self.noiselevel = 0.5
self.critic_loss_stat = WindowStat(500)
self.env_reward_stat = WindowStat(500)
self.shaping_reward_stat = WindowStat(500)
self.max_env_reward = 0
# thread to keep training
learn_thread = threading.Thread(target=self.keep_training)
learn_thread.setDaemon(True)
learn_thread.start()
self.create_actors()self.restore(args.restore_model_path)
# add lock between training and predicting
self.model_lock = threading.Lock()
# add lock when appending data to rpm or writing scalars to tensorboard
self.memory_lock = threading.Lock()
self.ready_actor_queue = queue.Queue()
self.total_steps = 0
self.noiselevel = 0.5
self.critic_loss_stat = WindowStat(500)
self.env_reward_stat = WindowStat(500)
self.shaping_reward_stat = WindowStat(500)
self.max_env_reward = 0
# thread to keep training
learn_thread = threading.Thread(target=self.keep_training)
learn_thread.setDaemon(True)
learn_thread.start()
self.create_actors()algorithm = parl.algorithms.A3C(
model, vf_loss_coeff=config['vf_loss_coeff'])
self.agent = AtariAgent(algorithm, config)
if machine_info.is_gpu_available():
assert get_gpu_count() == 1, 'Only support training in single GPU,\
Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_TO_USE]` .'
else:
cpu_num = os.environ.get('CPU_NUM')
assert cpu_num is not None and cpu_num == '1', 'Only support training in single CPU,\
Please set environment variable: `export CPU_NUM=1`.'
#========== Learner ==========
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.lr = None
self.entropy_coeff = None
self.learn_time_stat = TimeStat(100)
self.start_time = None
#========== Remote Actor ===========
self.remote_count = 0
self.sample_data_queue = queue.Queue()
self.remote_metrics_queue = queue.Queue()
self.sample_total_steps = 0self.rpm.load(args.restore_rpm_path)
if args.restore_model_path is not None:
self.restore(args.restore_model_path)
# add lock between training and predicting
self.model_lock = threading.Lock()
# add lock when appending data to rpm or writing scalars to tensorboard
self.memory_lock = threading.Lock()
self.ready_actor_queue = queue.Queue()
self.total_steps = 0
self.noiselevel = 0.5
self.critic_loss_stat = WindowStat(500)
self.env_reward_stat = WindowStat(500)
self.shaping_reward_stat = WindowStat(500)
self.max_env_reward = 0
# thread to keep training
learn_thread = threading.Thread(target=self.keep_training)
learn_thread.setDaemon(True)
learn_thread.start()
self.create_actors()self.params_lock = threading.Lock()
self.params_updated = False
self.cache_params_sent_cnt = 0
self.total_params_sync = 0
#========== Learner ==========
self.lr, self.entropy_coeff = None, None
self.lr_scheduler = PiecewiseScheduler(config['lr_scheduler'])
self.entropy_coeff_scheduler = PiecewiseScheduler(
config['entropy_coeff_scheduler'])
self.total_loss_stat = WindowStat(100)
self.pi_loss_stat = WindowStat(100)
self.vf_loss_stat = WindowStat(100)
self.entropy_stat = WindowStat(100)
self.kl_stat = WindowStat(100)
self.learn_time_stat = TimeStat(100)
self.start_time = None
self.learn_thread = threading.Thread(target=self.run_learn)
self.learn_thread.setDaemon(True)
self.learn_thread.start()
#========== Remote Actor ===========
self.remote_count = 0
self.batch_buffer = []
self.remote_metrics_queue = queue.Queue()
self.sample_total_steps = 0
self.create_actors()
Popular parl functions
- parl.connect
- parl.layers
- parl.layers.cast
- parl.layers.concat
- parl.layers.conv2d
- parl.layers.data
- parl.layers.fc
- parl.layers.squeeze
- parl.Model
- parl.remote.remote_constants
- parl.utils.logger
- parl.utils.logger.error
- parl.utils.logger.get_dir
- parl.utils.logger.info
- parl.utils.logger.warning
- parl.utils.machine_info.is_gpu_available
- parl.utils.tensorboard.add_scalar
- parl.utils.to_byte
- parl.utils.to_str
- parl.utils.window_stat.WindowStat