def __init__(self, base_name, observation_space, action_space, config):
common.a2c_common.ContinuousA2CBase.__init__(self, base_name,
observation_space,
action_space, config)
obs_shape = algos_torch.torch_ext.shape_whc_to_cwh(self.state_shape)
config = {
'actions_num': self.actions_num,
'input_shape': obs_shape,
'games_num': 1,
'batch_num': 1,
}
self.model = self.network.build(config)
self.model.cuda()
self.last_lr = float(self.last_lr)
self.optimizer = optim.Adam(self.model.parameters(),
float(self.last_lr))
#self.optimizer = algos_torch.torch_ext.RangerQH(self.model.parameters(), float(self.last_lr))
if self.normalize_input:
self.running_mean_std = RunningMeanStd(obs_shape).cuda()
if self.has_curiosity:
self.rnd_curiosity = rnd_curiosity.RNDCurisityTrain(
algos_torch.torch_ext.shape_whc_to_cwh(self.state_shape),
self.curiosity_config['network'], self.curiosity_config,
self.writer, lambda obs: self._preproc_obs(obs))
def __init__(self, config):
BasePlayer.__init__(self, config)
self.network = config['network']
self.actions_num = self.action_space.shape[0]
self.actions_low = self.action_space.low
self.actions_high = self.action_space.high
self.mask = [False]
observation_shape = algos_torch.torch_ext.shape_whc_to_cwh(
self.state_shape)
self.normalize_input = self.config['normalize_input']
print(self.state_shape)
config = {
'actions_num': self.actions_num,
'input_shape':
algos_torch.torch_ext.shape_whc_to_cwh(self.state_shape),
'games_num': 1,
'batch_num': 1,
}
self.model = self.network.build(config)
self.model.cuda()
self.model.eval()
if self.normalize_input:
self.running_mean_std = RunningMeanStd(observation_shape).cuda()
self.running_mean_std.eval()
def __init__(self, state_shape, model, config, writter, _preproc_obs):
nn.Module.__init__(self)
rnd_config = {
'input_shape': state_shape,
}
self.model = RNDCuriosityNetwork(model.build('rnd',
**rnd_config)).cuda()
self.config = config
self.lr = config['lr']
self.writter = writter
self.optimizer = torch.optim.Adam(self.model.parameters(),
float(self.lr))
self._preproc_obs = _preproc_obs
self.output_normalization = RunningMeanStd((1, ),
norm_only=True).cuda()
self.frame = 0
self.exp_percent = config.get('exp_percent', 1.0)
def __init__(self, base_name, observation_space, action_space, config):
common.a2c_common.DiscreteA2CBase.__init__(self, base_name,
observation_space,
action_space, config)
config = {
'actions_num': self.actions_num,
'input_shape':
algos_torch.torch_ext.shape_whc_to_cwh(self.state_shape),
'games_num': 1,
'batch_num': 1,
}
self.model = self.network.build(config)
self.model.cuda()
self.last_lr = float(self.last_lr)
self.optimizer = optim.Adam(self.model.parameters(),
float(self.last_lr))
#self.optimizer = algos_torch.torch_ext.RangerQH(self.model.parameters(), float(self.last_lr))
if self.normalize_input:
self.running_mean_std = RunningMeanStd(
observation_space.shape).cuda()