def __init__(self, env_args: Config, model_args: Config,
buffer_args: Config, train_args: Config):
# print("89898989")
self.env_args = env_args
self.model_args = model_args
self.buffer_args = buffer_args
self.train_args = train_args
self.use_GCN = False
self.model_index = str(self.train_args.get('index'))
self.all_learner_print = bool(
self.train_args.get('all_learner_print', False))
if '-' not in self.train_args['name']:
self.train_args['name'] += f'-{self.model_index}'
if self.model_args['load'] is None:
self.train_args['load_model_path'] = os.path.join(
self.train_args['base_dir'], self.train_args['name'])
else:
if '/' in self.model_args['load'] or '\\' in self.model_args[
'load']: # 所有训练进程都以该模型路径初始化，绝对路径
self.train_args['load_model_path'] = self.model_args['load']
elif '-' in self.model_args['load']:
self.train_args['load_model_path'] = os.path.join(
self.train_args['base_dir'],
self.model_args['load']) # 指定了名称和序号，所有训练进程都以该模型路径初始化，相对路径
else: # 只写load的训练名称，不用带进程序号，会自动补
self.train_args['load_model_path'] = os.path.join(
self.train_args['base_dir'],
self.model_args['load'] + f'-{self.model_index}')
# ENV
self.env = make_env(self.env_args.to_dict, self.use_GCN)
# ALGORITHM CONFIG
Model, algorithm_config, _policy_mode = get_model_info(
self.model_args['algo'])
self.model_args['policy_mode'] = _policy_mode
if self.model_args['algo_config'] is not None:
algorithm_config = UpdateConfig(algorithm_config,
self.model_args['algo_config'],
'algo')
ShowConfig(algorithm_config)
# BUFFER
if _policy_mode == 'off-policy':
self.buffer_args['batch_size'] = algorithm_config['batch_size']
self.buffer_args['buffer_size'] = algorithm_config['buffer_size']
if self.model_args['algo'] in ['drqn', 'drdqn']:
self.buffer_args['type'] = 'EpisodeER'
else:
_use_priority = algorithm_config.get('use_priority', False)
_n_step = algorithm_config.get('n_step', False)
if _use_priority and _n_step:
self.buffer_args['type'] = 'NstepPER'
self.buffer_args['NstepPER'][
'max_episode'] = self.train_args['max_episode']
self.buffer_args['NstepPER']['gamma'] = algorithm_config[
'gamma']
algorithm_config['gamma'] = pow(
algorithm_config['gamma'], self.buffer_args['NstepPER']
['n']) # update gamma for n-step training.
elif _use_priority:
self.buffer_args['type'] = 'PER'
self.buffer_args['PER']['max_episode'] = self.train_args[
'max_episode']
elif _n_step:
self.buffer_args['type'] = 'NstepER'
self.buffer_args['NstepER']['gamma'] = algorithm_config[
'gamma']
algorithm_config['gamma'] = pow(
algorithm_config['gamma'],
self.buffer_args['NstepER']['n'])
else:
self.buffer_args['type'] = 'ER'
else:
self.buffer_args['type'] = 'Pandas'
# MODEL
base_dir = os.path.join(
self.train_args['base_dir'], self.train_args['name']
) # train_args['base_dir'] DIR/ENV_NAME/ALGORITHM_NAME
if 'batch_size' in algorithm_config.keys() and train_args['fill_in']:
self.train_args['pre_fill_steps'] = algorithm_config['batch_size']
if self.env_args['type'] == 'gym':
self.eval_env_args = deepcopy(self.env_args)
self.eval_env_args.env_num = 1
self.eval_env = make_env(self.eval_env_args.to_dict)
# buffer ------------------------------
if 'Nstep' in self.buffer_args[
'type'] or 'Episode' in self.buffer_args['type']:
self.buffer_args[self.buffer_args['type']][
'agents_num'] = self.env_args['env_num']
self.buffer = get_buffer(self.buffer_args)
# buffer ------------------------------
# model -------------------------------
model_params = {
's_dim': self.env.s_dim,
'visual_sources': self.env.visual_sources,
'visual_resolution': self.env.visual_resolution,
'a_dim_or_list': self.env.a_dim_or_list,
'is_continuous': self.env.is_continuous,
'max_episode': self.train_args.max_episode,
'base_dir': base_dir,
'logger2file': self.model_args.logger2file,
'seed': self.model_args.seed
}
self.model = Model(**model_params, **algorithm_config)
self.model.set_buffer(self.buffer)
self.model.init_or_restore(self.train_args['load_model_path'])
# model -------------------------------
self.train_args['begin_episode'] = self.model.get_init_episode()
if not self.train_args['inference']:
records_dict = {
'env': self.env_args.to_dict,
'model': self.model_args.to_dict,
'buffer': self.buffer_args.to_dict,
'train': self.train_args.to_dict,
'algo': algorithm_config
}
save_config(os.path.join(base_dir, 'config'), records_dict)
else:
# buffer -----------------------------------
self.buffer_args_s = []
for i in range(self.env.brain_num):
_bargs = deepcopy(self.buffer_args)
if 'Nstep' in _bargs['type'] or 'Episode' in _bargs['type']:
_bargs[_bargs['type']][
'agents_num'] = self.env.brain_agents[i]
self.buffer_args_s.append(_bargs)
buffers = [
get_buffer(self.buffer_args_s[i])
for i in range(self.env.brain_num)
]
# buffer -----------------------------------
# model ------------------------------------
self.model_args_s = []
for i in range(self.env.brain_num):
_margs = deepcopy(self.model_args)
_margs['seed'] = self.model_args['seed'] + i * 10
self.model_args_s.append(_margs)
model_params = [
{
's_dim': self.env.s_dim[i],
'a_dim_or_list': self.env.a_dim_or_list[i],
'visual_sources': self.env.visual_sources[i],
'visual_resolution': self.env.visual_resolutions[i],
'is_continuous': self.env.is_continuous[i],
'max_episode': self.train_args.max_episode,
'base_dir': os.path.join(base_dir, b),
'logger2file': self.model_args_s[i].logger2file,
'seed': self.model_args_s[i].
seed, # 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100
} for i, b in enumerate(self.env.brain_names)
]
# multi agent training------------------------------------
if self.model_args['algo'][:3] == 'ma_':
self.ma = True
assert self.env.brain_num > 1, 'if using ma* algorithms, number of brains must larger than 1'
self.ma_data = ExperienceReplay(batch_size=10, capacity=1000)
[
mp.update({
'n': self.env.brain_num,
'i': i
}) for i, mp in enumerate(model_params)
]
else:
self.ma = False
# multi agent training------------------------------------
self.models = [
Model(**model_params[i], **algorithm_config)
for i in range(self.env.brain_num)
]
[
model.set_buffer(buffer)
for model, buffer in zip(self.models, buffers)
]
[
self.models[i].init_or_restore(
os.path.join(self.train_args['load_model_path'], b))
for i, b in enumerate(self.env.brain_names)
]
# model ------------------------------------
self.train_args['begin_episode'] = self.models[0].get_init_episode(
)
if not self.train_args['inference']:
for i, b in enumerate(self.env.brain_names):
records_dict = {
'env': self.env_args.to_dict,
'model': self.model_args_s[i].to_dict,
'buffer': self.buffer_args_s[i].to_dict,
'train': self.train_args.to_dict,
'algo': algorithm_config
}
save_config(os.path.join(base_dir, b, 'config'),
records_dict)
def gym_run(default_args, share_args, options, max_step, max_episode,
save_frequency, name):
from gym_loop import Loop
from gym.spaces import Box, Discrete, Tuple
from gym_wrapper import gym_envs
try:
tf_version, (model, policy_mode,
_) = get_model_info(options['--algorithm'])
algorithm_config = sth.load_config(
f'./Algorithms/{tf_version}/config.yaml')[options['--algorithm']]
except KeyError:
raise NotImplementedError
available_type = [Box, Discrete]
render_episode = int(options['--render-episode']) if options[
'--render-episode'] != 'None' else default_args['render_episode']
try:
env = gym_envs(options['--gym-env'], int(options['--gym-agents']))
assert type(env.observation_space) in available_type and type(
env.action_space
) in available_type, 'action_space and observation_space must be one of available_type'
except Exception as e:
print(e)
if options['--config-file'] != 'None':
algorithm_config = update_config(algorithm_config,
options['--config-file'])
_base_dir = os.path.join(share_args['base_dir'], options['--gym-env'],
options['--algorithm'])
base_dir = os.path.join(_base_dir, name)
show_config(algorithm_config)
if type(env.observation_space) == Box:
s_dim = env.observation_space.shape[0] if len(
env.observation_space.shape) == 1 else 0
else:
s_dim = int(env.observation_space.n)
if len(env.observation_space.shape) == 3:
visual_sources = 1
visual_resolution = list(env.observation_space.shape)
else:
visual_sources = 0
visual_resolution = []
if type(env.action_space) == Box:
assert len(
env.action_space.shape
) == 1, 'if action space is continuous, the shape length of action must equal to 1'
a_dim_or_list = env.action_space.shape
action_type = 'continuous'
elif type(env.action_space) == Tuple:
assert all(
[type(i) == Discrete for i in env.action_space]
) == True, 'if action space is Tuple, each item in it must have type Discrete'
a_dim_or_list = [i.n for i in env.action_space]
action_type = 'discrete'
else:
a_dim_or_list = [env.action_space.n]
action_type = 'discrete'
gym_model = model(s_dim=s_dim,
visual_sources=visual_sources,
visual_resolution=visual_resolution,
a_dim_or_list=a_dim_or_list,
action_type=action_type,
max_episode=max_episode,
base_dir=base_dir,
logger2file=share_args['logger2file'],
out_graph=share_args['out_graph'],
**algorithm_config)
gym_model.init_or_restore(
os.path.join(
_base_dir,
name if options['--load'] == 'None' else options['--load']))
begin_episode = gym_model.get_init_episode()
params = {
'env': env,
'gym_model': gym_model,
'action_type': action_type,
'begin_episode': begin_episode,
'save_frequency': save_frequency,
'max_step': max_step,
'max_episode': max_episode,
'eval_while_train':
default_args['eval_while_train'], # whether to eval while training.
'max_eval_episode': default_args['max_eval_episode'],
'render': default_args['render'],
'render_episode': render_episode,
'policy_mode': policy_mode
}
if 'batch_size' in algorithm_config.keys() and options['--fill-in']:
steps = algorithm_config['batch_size']
else:
steps = default_args['random_steps']
if options['--inference']:
Loop.inference(env, gym_model, action_type)
else:
sth.save_config(os.path.join(base_dir, 'config'), algorithm_config)
try:
Loop.no_op(env,
gym_model,
action_type,
steps,
choose=options['--noop-choose'])
Loop.train(**params)
except Exception as e:
print(e)
finally:
try:
gym_model.close()
except Exception as e:
print(e)
finally:
env.close()
sys.exit()
def unity_run(default_args, share_args, options, max_step, max_episode,
save_frequency, name):
from mlagents.envs import UnityEnvironment
from utils.sampler import create_sampler_manager
try:
tf_version, (model, policy_mode,
_) = get_model_info(options['--algorithm'])
algorithm_config = sth.load_config(
f'./Algorithms/{tf_version}/config.yaml')[options['--algorithm']]
ma = options['--algorithm'][:3] == 'ma_'
except KeyError:
raise NotImplementedError
reset_config = default_args['reset_config']
if options['--unity']:
env = UnityEnvironment()
env_name = 'unity'
else:
file_name = default_args['exe_file'] if options[
'--env'] == 'None' else options['--env']
if os.path.exists(file_name):
env = UnityEnvironment(file_name=file_name,
base_port=int(options['--port']),
no_graphics=False if options['--inference']
else not options['--graphic'])
env_dir = os.path.split(file_name)[0]
env_name = os.path.join(*env_dir.replace('\\', '/').replace(
r'//', r'/').split('/')[-2:])
sys.path.append(env_dir)
if os.path.exists(env_dir + '/env_config.py'):
import env_config
reset_config = env_config.reset_config
max_step = env_config.max_step
if os.path.exists(env_dir + '/env_loop.py'):
from env_loop import Loop
else:
raise Exception('can not find this file.')
sampler_manager, resampling_interval = create_sampler_manager(
options['--sampler'], env.reset_parameters)
if 'Loop' not in locals().keys():
if ma:
from ma_loop import Loop
else:
from loop import Loop
if options['--config-file'] != 'None':
algorithm_config = update_config(algorithm_config,
options['--config-file'])
_base_dir = os.path.join(share_args['base_dir'], env_name,
options['--algorithm'])
base_dir = os.path.join(_base_dir, name)
show_config(algorithm_config)
brain_names = env.external_brain_names
brains = env.brains
brain_num = len(brain_names)
visual_resolutions = {}
for i in brain_names:
if brains[i].number_visual_observations:
visual_resolutions[f'{i}'] = [
brains[i].camera_resolutions[0]['height'],
brains[i].camera_resolutions[0]['width'],
1 if brains[i].camera_resolutions[0]['blackAndWhite'] else 3
]
else:
visual_resolutions[f'{i}'] = []
model_params = [{
's_dim':
brains[i].vector_observation_space_size *
brains[i].num_stacked_vector_observations,
'a_dim_or_list':
brains[i].vector_action_space_size,
'action_type':
brains[i].vector_action_space_type,
'max_episode':
max_episode,
'base_dir':
os.path.join(base_dir, i),
'logger2file':
share_args['logger2file'],
'out_graph':
share_args['out_graph'],
} for i in brain_names]
if ma:
assert brain_num > 1, 'if using ma* algorithms, number of brains must larger than 1'
data = ExperienceReplay(share_args['ma']['batch_size'],
share_args['ma']['capacity'])
extra_params = {'data': data}
models = [
model(n=brain_num, i=i, **model_params[i], **algorithm_config)
for i in range(brain_num)
]
else:
extra_params = {}
models = [
model(visual_sources=brains[i].number_visual_observations,
visual_resolution=visual_resolutions[f'{i}'],
**model_params[index],
**algorithm_config) for index, i in enumerate(brain_names)
]
[
models[index].init_or_restore(
os.path.join(
_base_dir,
name if options['--load'] == 'None' else options['--load'], i))
for index, i in enumerate(brain_names)
]
begin_episode = models[0].get_init_episode()
params = {
'env': env,
'brain_names': brain_names,
'models': models,
'begin_episode': begin_episode,
'save_frequency': save_frequency,
'reset_config': reset_config,
'max_step': max_step,
'max_episode': max_episode,
'sampler_manager': sampler_manager,
'resampling_interval': resampling_interval,
'policy_mode': policy_mode
}
if 'batch_size' in algorithm_config.keys() and options['--fill-in']:
steps = algorithm_config['batch_size']
else:
steps = default_args['no_op_steps']
no_op_params = {
'env': env,
'brain_names': brain_names,
'models': models,
'brains': brains,
'steps': steps,
'choose': options['--noop-choose']
}
params.update(extra_params)
no_op_params.update(extra_params)
if options['--inference']:
Loop.inference(env,
brain_names,
models,
reset_config=reset_config,
sampler_manager=sampler_manager,
resampling_interval=resampling_interval)
else:
try:
[
sth.save_config(os.path.join(base_dir, i, 'config'),
algorithm_config) for i in brain_names
]
Loop.no_op(**no_op_params)
Loop.train(**params)
except Exception as e:
print(e)
finally:
try:
[models[i].close() for i in range(len(models))]
except Exception as e:
print(e)
finally:
env.close()
sys.exit()
def gym_run(default_args, share_args, options, max_step, max_episode, save_frequency, name, seed):
from gym_loop import Loop
from gym_wrapper import gym_envs
model, algorithm_config, policy_mode = get_model_info(options['--algorithm'])
render_episode = int(options['--render-episode']) if options['--render-episode'] != 'None' else default_args['render_episode']
try:
env = gym_envs(gym_env_name=options['--gym-env'],
n=int(options['--gym-agents']),
seed=int(options['--gym-env-seed']),
render_mode=default_args['render_mode'])
except Exception as e:
print(e)
if options['--config-file'] != 'None':
algorithm_config = update_config(algorithm_config, options['--config-file'])
_base_dir = os.path.join(share_args['base_dir'], options['--gym-env'], options['--algorithm'])
base_dir = os.path.join(_base_dir, name)
show_config(algorithm_config)
model_params = {
's_dim': env.s_dim,
'visual_sources': env.visual_sources,
'visual_resolution': env.visual_resolution,
'a_dim_or_list': env.a_dim_or_list,
'is_continuous': env.is_continuous,
'max_episode': max_episode,
'base_dir': base_dir,
'logger2file': share_args['logger2file'],
'seed': seed,
}
gym_model = model(
**model_params,
**algorithm_config
)
gym_model.init_or_restore(os.path.join(_base_dir, name if options['--load'] == 'None' else options['--load']))
begin_episode = gym_model.get_init_episode()
params = {
'env': env,
'gym_model': gym_model,
'begin_episode': begin_episode,
'save_frequency': save_frequency,
'max_step': max_step,
'max_episode': max_episode,
'eval_while_train': default_args['eval_while_train'], # whether to eval while training.
'max_eval_episode': default_args['max_eval_episode'],
'render': default_args['render'],
'render_episode': render_episode,
'policy_mode': policy_mode
}
if 'batch_size' in algorithm_config.keys() and options['--fill-in']:
steps = algorithm_config['batch_size']
else:
steps = default_args['random_steps']
if options['--inference']:
Loop.inference(env, gym_model)
else:
sth.save_config(os.path.join(base_dir, 'config'), algorithm_config)
try:
Loop.no_op(env, gym_model, steps, choose=options['--noop-choose'])
Loop.train(**params)
except Exception as e:
print(e)
finally:
gym_model.close()
env.close()
sys.exit()