def __init__(self,action_dim,state_dim,agentParam,useLaw,useCenCritc,num_agent,CNN=False, width=None, height=None, channel=None):
self.CNN = CNN
self.device = agentParam["device"]
if CNN:
self.CNN_preprocessA = CNN_preprocess(width,height,channel)
self.CNN_preprocessC = CNN_preprocess(width,height,channel)
state_dim = self.CNN_preprocessA.get_state_dim()
#if agentParam["ifload"]:
#self.actor = torch.load(agentParam["filename"]+"actor_"+agentParam["id"]+".pth",map_location = torch.device('cuda'))
#self.critic = torch.load(agentParam["filename"]+"critic_"+agentParam["id"]+".pth",map_location = torch.device('cuda'))
#else:
if useLaw:
self.actor = ActorLaw(action_dim,state_dim).to(self.device)
else:
self.actor = Actor(action_dim,state_dim).to(self.device)
if useCenCritc:
self.critic = Centralised_Critic(state_dim,num_agent).to(self.device)
else:
self.critic = Critic(state_dim).to(self.device)
self.action_dim = action_dim
self.state_dim = state_dim
self.noise_epsilon = 0.99
self.constant_decay = 0.1
self.optimizerA = torch.optim.Adam(self.actor.parameters(), lr = 0.001)
self.optimizerC = torch.optim.Adam(self.critic.parameters(), lr = 0.001)
self.lr_scheduler = {"optA":torch.optim.lr_scheduler.StepLR(self.optimizerA,step_size=1000,gamma=0.9,last_epoch=-1),
"optC":torch.optim.lr_scheduler.StepLR(self.optimizerC,step_size=1000,gamma=0.9,last_epoch=-1)}
if CNN:
# self.CNN_preprocessA = CNN_preprocess(width,height,channel)
# self.CNN_preprocessC = CNN_preprocess
self.optimizerA = torch.optim.Adam(itertools.chain(self.CNN_preprocessA.parameters(),self.actor.parameters()),lr=0.0001)
self.optimizerC = torch.optim.Adam(itertools.chain(self.CNN_preprocessC.parameters(),self.critic.parameters()),lr=0.001)
self.lr_scheduler = {"optA": torch.optim.lr_scheduler.StepLR(self.optimizerA, step_size=10000, gamma=0.9, last_epoch=-1),
"optC": torch.optim.lr_scheduler.StepLR(self.optimizerC, step_size=10000, gamma=0.9, last_epoch=-1)}
def __init__(self,
action_dim,
state_dim,
CNN=False,
width=None,
height=None,
channel=None,
device='cpu'):
self.CNN = CNN
if CNN:
self.CNN_preprocessA = CNN_preprocess(width, height, channel)
self.CNN_preprocessC = CNN_preprocess(width, height, channel)
state_dim = self.CNN_preprocessA.get_state_dim()
self.device = device
self.actor = Actor(action_dim, state_dim)
self.critic = Critic(state_dim)
self.action_dim = action_dim
self.state_dim = state_dim
self.noise_epsilon = 0.999
self.constant_decay = 1
self.optimizerA = torch.optim.Adam(self.actor.parameters(), lr=0.00001)
self.optimizerC = torch.optim.Adam(self.critic.parameters(), lr=0.01)
self.lr_scheduler = {
"optA":
torch.optim.lr_scheduler.StepLR(self.optimizerA,
step_size=1000,
gamma=1,
last_epoch=-1),
"optC":
torch.optim.lr_scheduler.StepLR(self.optimizerC,
step_size=1000,
gamma=0.9,
last_epoch=-1)
}
if CNN:
# self.CNN_preprocessA = CNN_preprocess(width,height,channel)
# self.CNN_preprocessC = CNN_preprocess
self.optimizerA = torch.optim.Adam(itertools.chain(
self.CNN_preprocessA.parameters(), self.actor.parameters()),
lr=0.0001)
self.optimizerC = torch.optim.Adam(itertools.chain(
self.CNN_preprocessC.parameters(), self.critic.parameters()),
lr=0.001)
self.lr_scheduler = {
"optA":
torch.optim.lr_scheduler.StepLR(self.optimizerA,
step_size=10000,
gamma=1,
last_epoch=-1),
"optC":
torch.optim.lr_scheduler.StepLR(self.optimizerC,
step_size=10000,
gamma=0.9,
last_epoch=-1)
}