class HIRO(object):
def __init__(self, params, name, task):
self.name = name
self.task = task
self.vPars = params['valPars']
self.vTrain = params['valTrain']
self.mPars = params['mPars']
self.mTrain = params['mTrain']
self.wPars = params['actPars']
self.wTrain = params['actTrain']
self.w_vPars = params['w_vPars']
self.w_vTrain = params['w_vTrain']
self.agents = params['agents']
self.pubs = {}
for key in self.agents.keys():
bot = self.agents[key]
self.pubs[key] = rospy.Publisher(bot['pub'], Vector3, queue_size = 1)
rospy.Subscriber("/finished", Int8, self.receiveDone, queue_size = 1)
self.valueLoss = []
self.manager = Network(self.mPars, self.mTrain)
self.m_critic = Network(self.vPars, self.vTrain)
self.m_critic_target= Network(self.vPars, self.vTrain)
self.worker = Network(self.wPars, self.wTrain)
self.w_critic = Network(self.w_vPars, self.w_vTrain)
self.w_critic_target= Network(self.w_vPars, self.w_vTrain)
self.m_discount = self.vTrain['m_gamma']
self.w_discount = self.vTrain['w_gamma']
self.lr = self.vTrain['lr']
self.trainMode = self.vPars['trainMode']
self.step = self.vTrain['step']
self.stop = False
self.c = self.mTrain['c']
self.tau = .005
self.noise = Noise(self.manager.neurons[-1], theta = .4, max_sigma = .2, min_sigma = 0, decay = 1)
self.exp = Memory()
self.temp = []
self.totalSteps = 0
self.soft = nn.Softmax(dim=1)
self.reset()
task.initAgent(self)
while(not self.stop):
x = 1+1
task.postTraining()
def receiveDone(self, message):
if message.data == 1: #all iterations are done. Check manager.py
self.stop = True
if message.data == 2: #timed out. Check manager.py
self.task.restartProtocol(restart = 1)
def get_action(self, s, s_w = None):
s = torch.FloatTensor(s)
if self.iteration % self.c == 0:
self.goal = self.manager(s)
noise = torch.FloatTensor(self.noise.get_noise())
self.goal += noise
else:
self.goal = self.prevState[:,:2] + self.goal - s[:,:2]
self.temp_second = self.temp_first
self.temp_first = self.goal
self.prevState = s
s = s[:,:6]
inpt = torch.cat((s, self.goal), dim=1)
policy = self.worker(inpt)
policy = self.soft(policy)
choice = np.asscalar(self.choose(policy))
self.iteration += 1
return choice #single env
def choose(self, policies):
m = Categorical(policies)
action = m.sample()
action = action.data.cpu().numpy()
return action
def saveModel(self):
pass
def store(self, s, a, r, sprime, aprime, done):
if self.temp_second != None:
self.temp.append(Transition(s, a, r, 1-done, sprime, None, self.temp_second.detach().numpy(), self.goal.detach().numpy()))
if self.iteration % self.c == 1 and self.iteration != 1: # remember, we push at 1 because we incremented in get_action
self.temp = Transition(*zip(*self.temp))
self.exp.push(self.temp) # store into exp
self.temp = []
def reset(self):
self.iteration = 0
self.temp_first, self.temp_second = (None, None)
self.prevState = None
self.temp = []
return
def generateSamples(self, goals, states, next_states):
next_states = next_states[:, :2]
states = states[:, :2]
candidates = (next_states - states).unsqueeze(0)
candidates = torch.cat((candidates, goals.unsqueeze(0)), dim=0)
normal = Normal(next_states - states, torch.ones(next_states.size()) / 2)
sampled = normal.sample((8,))
candidates = torch.cat((candidates, sampled), dim=0)
# return shape (# candidates, batch_size, dimensions of goal)
return candidates
def getTransitions(self, initial_goals, states, next_states):
# initial_goals shape: (# candidates ,batch_size, dimensions of goal)
# states shape: (batch_size, c, dimensions of state)
states = states[:,:, :2]
next_states = next_states[:,:,:2]
goals = [initial_goals.unsqueeze(0)]
for c in range(self.c - 1):
prev = goals[-1].squeeze(0)
curr = states[:, c, :] + prev - next_states[:,c,:] # broadcast. This should take shape of initial_goals
goals.append(curr.unsqueeze(0))
goals = torch.cat(goals, dim=0)
goals = goals.transpose(0,1)
goals = goals.transpose(1,2)
# return shape (# candidates, batch_size, c, dimensions of goal)
return goals
def getProbabilities(self, transitions, states, actions):
# transitions shape (# candidates, batch_size, c, dimensions of goal)
# states shape: (batch_size, c, dimensions of state)
# actions shape: (batch_size, c)
states = states[:, :, :6]
states = states.unsqueeze(0)
size = states.size()
states = states.expand(transitions.size()[0], size[1], size[2], size[3])
inpt = torch.cat((states, transitions), dim=3)
soft = nn.Softmax(dim = 3)
actions = actions.expand(transitions.size()[0], actions.size()[0], actions.size()[1]).unsqueeze(3)
probs = soft(self.worker(inpt)).gather(3, actions.long()).squeeze(3)
probs = torch.prod(probs, dim=2)
# return shape (# candidates, batch_size) of probabilities
return probs
def train(self):
if len(self.exp) > 300:
groups = self.exp.sample(self.step) # sample groupings of samples
m_states = torch.cat(map(lambda g: torch.Tensor(g.state[0]), groups), dim=0)
m_next_states = torch.cat(map(lambda g: torch.Tensor(g.next_state[-1]), groups), dim=0)
m_goals = torch.cat(map(lambda g: torch.Tensor(g.goal[0]), groups), dim=0)
m_rewards = torch.Tensor(map(lambda g: sum(g.reward), groups)).squeeze(2)
m_masks = torch.Tensor(map(lambda g: g.mask[-1], groups)).unsqueeze(1)
w_states = torch.cat(map(lambda g: torch.Tensor(g.state).unsqueeze(0), groups), dim=0).squeeze()
w_next_states = torch.cat(map(lambda g: torch.Tensor(g.next_state).unsqueeze(0), groups), dim=0).squeeze()
w_actions = torch.cat(map(lambda g: torch.Tensor(g.action).unsqueeze(0), groups), dim=0)
candidates = self.generateSamples(m_goals, m_states, m_next_states)
cand_transitions = self.getTransitions(candidates, w_states, w_next_states)
probs = self.getProbabilities(cand_transitions, w_states, w_actions)
cand_indices = probs.argmax(dim=0).unsqueeze(0).unsqueeze(2)
cand_indices = cand_indices.expand(cand_indices.size()[0], cand_indices.size()[1], candidates.size()[2])
m_goals = candidates.gather(0, cand_indices).squeeze() #size: (batch_size, dimension of goals)
states = []
actions = []
next_states = []
masks = []
goals = []
next_goals = []
for g in groups:
states.append(torch.Tensor(g.state).squeeze()[:, :6])
actions.append(torch.Tensor(g.action).squeeze())
next_states.append(torch.Tensor(g.next_state).squeeze()[:, :6])
masks.append(torch.Tensor(g.mask).squeeze())
goals.append(torch.Tensor(g.goal).squeeze())
next_goals.append(torch.Tensor(g.next_goal).squeeze())
states = torch.cat(states, dim=0)
actions = torch.cat(actions, dim=0).unsqueeze(1)
next_states = torch.cat(next_states, dim=0)
masks = torch.cat(masks, dim=0).unsqueeze(1)
goals = torch.cat(goals, dim=0)
next_goals = torch.cat(next_goals, dim=0)
rewards = -torch.norm(states[:,:2] + goals - next_states[:,:2], dim=1).unsqueeze(1)
# Manager critic
q = self.m_critic(torch.cat((m_states, m_goals), dim=1))
m_next_actions = self.manager(m_next_states)
q_tar = m_rewards + self.m_discount * self.m_critic_target(torch.cat((m_next_states, m_next_actions), dim=1))
loss = self.m_critic.get_loss(q, q_tar.detach())
self.m_critic.optimizer.zero_grad()
loss.backward()
self.m_critic.optimizer.step()
# Manager actor
new_actions = self.manager(m_states)
q = self.m_critic(torch.cat((m_states, new_actions), dim=1))
loss = -q.mean()
self.manager.optimizer.zero_grad()
loss.backward()
self.m_critic.optimizer.step()
# Worker critic
q = self.w_critic(torch.cat((states, goals), dim=1)).gather(1, actions.long())
next_actions = self.worker(torch.cat((next_states, next_goals), dim=1))
next_actions = self.choose(self.soft(next_actions))
q_tar = rewards + self.w_discount * masks * self.w_critic_target(torch.cat((next_states, next_goals), dim=1)).gather(1, torch.Tensor(next_actions).long().unsqueeze(1))
loss = self.w_critic.get_loss(q, q_tar.detach())
self.w_critic.optimizer.zero_grad()
loss.backward()
self.w_critic.optimizer.step()
# Worker actor
new_actions = self.worker(torch.cat((states[:,:6], goals), dim=1))
policy = self.soft(new_actions)
new_actions = self.choose(policy)
q = self.w_critic(torch.cat((states, goals), dim=1))
q = q.gather(1, torch.Tensor(new_actions).long().unsqueeze(1))
loss = -q.mean()
self.worker.optimizer.zero_grad()
loss.backward()
self.worker.optimizer.step()
for target_param, param in zip(self.m_critic_target.parameters(), self.m_critic.parameters()):
target_param.data.copy_(target_param.data * (1.0 - self.tau) + param.data * self.tau)
for target_param, param in zip(self.w_critic_target.parameters(), self.w_critic.parameters()):
target_param.data.copy_(target_param.data * (1.0 - self.tau) + param.data * self.tau)
# Push updated replay entires into replay
for i, goal in enumerate(m_goals):
curr_group = groups[i]
curr_goal = goal.unsqueeze(0).detach().numpy()
inserts = (curr_goal)
for j in range(self.c - 1):
curr_goal = curr_group.state[j][:,:2].reshape(1,-1) + curr_goal - curr_group.next_state[j][:,:2].reshape(1,-1)
inserts = inserts + (curr_goal)
curr_group._replace(goal=inserts)
self.exp.push(curr_group)
self.totalSteps += 1
return loss
class CounterContinuous(object):
def __init__(self, params, name, task):
self.name = name
self.task = task
self.vTrain = params['valTrain']
self.vPars = params['valPars']
self.aTrain = params['actTrain']
self.aPars = params['actPars']
self.agents = params['agents']
self.pubs = {}
for key in self.agents.keys():
bot = self.agents[key]
self.pubs[key] = rospy.Publisher(bot['pub'], Vector3, queue_size=1)
rospy.Subscriber("/finished", Int8, self.receiveDone, queue_size=1)
self.valueLoss = []
self.actorLoss = []
self.h_state_n = self.aPars['h_state_n']
self.x_state_n = self.aPars['x_state_n']
self.u_n = self.aPars['u_n']
self.clip_grad_norm = self.aTrain['clip']
self.homogenous = self.aPars['share_params']
self.critic = Network(self.vPars, self.vTrain).to(device)
self.target = Network(self.vPars, self.vTrain).to(device)
if self.homogenous:
self.actor = SoftNetwork(self.aPars, self.aTrain).to(device)
else:
self.actor = [
SoftNetwork(self.aPars, self.aTrain)
for i in range(len(self.agents))
]
for target_param, param in zip(self.target.parameters(),
self.critic.parameters()):
target_param.data.copy_(param.data)
self.clip_grad_norm = self.aTrain['clip']
self.trainMode = self.vPars['trainMode']
self.batch_size = self.vTrain['batch_size']
self.discount = self.vTrain['gamma']
self.range = self.aPars['mean_range']
self.td_lambda = .8
self.tau = .005
self.lower_bound = self.aTrain['clamp'][2]
self.stop = False
self.trained = False
self.exp = Memory()
self.totalSteps = 0
self.reset()
task.initAgent(self)
while (not self.stop):
x = 1 + 1
task.postTraining()
def receiveDone(self, message):
if message.data == 1: #all iterations are done. Check manager.py
self.stop = True
if message.data == 2: #timed out. Check manager.py
self.task.restartProtocol(restart=1)
def get_action(self, s_true, s_split):
if self.homogenous:
a1, log_prob1, z, mu1, log_std1 = self.actor(
torch.FloatTensor(s_split[0]))
a2, log_prob2, z, mu2, log_std2 = self.actor(
torch.FloatTensor(s_split[1]))
else: # TODO: Fix this below:
a1, h_new1, log_prob1, mu1, std1 = self.actor[0](torch.FloatTensor(
s_split[0]), self.h[0])
a2, h_new2, log_prob2, mu2, std2 = self.actor[1](torch.FloatTensor(
s_split[1]), self.h[1])
action = [a1.detach().numpy().ravel(), a2.detach().numpy().ravel()]
return [a1, a2]
def choose(self, policies):
m = Categorical(policies)
action = m.sample()
action = action.data.cpu().numpy()
return np.asscalar(action)
def saveModel(self):
pass
def store(self, s, a, r, sprime, aprime, done, local, next_local):
self.exp.push(s, a, r, 1 - done, aprime, sprime, local, next_local)
def reset(self):
curr = self.actor.clamp[0]
if self.trained:
new = max(self.lower_bound, .05 * self.lower_bound + .95 * curr)
self.actor.clamp = (new, self.actor.clamp[1], self.lower_bound)
self.trained = False
return
def get_grad_norm(self, model):
total_norm = 0
for p in model.parameters():
if p.grad is None:
continue
param_norm = p.grad.data.norm(2)
total_norm += param_norm.item()**2
grad_norm = total_norm**(1. / 2)
return grad_norm
def get_lambda_targets(self, rewards, mask, gamma, target_qs):
target_qs = target_qs.squeeze()
ret = target_qs.new_zeros(*target_qs.shape)
ret[-1] = target_qs[-1] * mask[-1]
for t in range(ret.size()[0] - 2, -1, -1):
ret[t] = self.td_lambda * gamma * ret[t + 1] + \
mask[t] * (rewards[t] + (1 - self.td_lambda) * gamma * target_qs[t + 1])
return ret.unsqueeze(1)
def zipStack(self, data):
data = zip(*data)
data = [torch.stack(d).squeeze().to(device) for d in data]
return data
def monte_carlo(self, mean, std, n=500):
# returns tensors representing n sampled from mean and std
normal = Normal(mean, std)
return normal.sample((n, ))
def train(self, episode_done=False):
if len(self.exp) >= 500:
transition = self.exp.sample(self.batch_size)
states = torch.squeeze(torch.Tensor(transition.state)).to(device)
actions = self.zipStack(transition.action)
rewards = torch.Tensor(transition.reward).to(device)
states_next = torch.squeeze(torch.Tensor(
transition.next_state)).to(device)
masks = torch.Tensor(transition.mask).to(device)
local = self.zipStack(transition.local)
next_local = self.zipStack(transition.next_local)
actions_next = []
for s in next_local:
a, log_prob, _, _, _ = self.actor(s)
actions_next.append(a.detach())
inp = torch.cat((states_next, actions_next[0], actions_next[1]),
dim=1)
q_tar = rewards.unsqueeze(
1) + self.discount * masks.unsqueeze(1) * self.target(inp)
inp = torch.cat((states, actions[0].detach(), actions[1].detach()),
dim=1)
q = self.critic(inp)
loss = self.critic.get_loss(q, q_tar.detach())
self.critic.optimizer.zero_grad()
loss.backward()
self.critic.optimizer.step()
self.valueLoss.append(loss)
actor_loss = 0
actions = []
means = []
log_stds = []
log_probs = []
for s in local:
a, log_prob, z, mu, log_std = self.actor(s)
actions.append(a)
means.append(mu)
log_stds.append(log_std)
log_probs.append(log_prob)
# train first agent
inp = torch.cat((states, actions[0], actions[1].detach()), dim=1)
q_out = self.critic(inp)
samples = self.monte_carlo(means[0], log_stds[0].exp())
samples = self.range * torch.tanh(samples)
repeat_s = states.unsqueeze(0)
repeat_s = repeat_s.expand(samples.size()[0],
repeat_s.size()[1],
repeat_s.size()[2])
repeat_a = actions[1].unsqueeze(0)
repeat_a = repeat_a.expand(samples.size()[0],
repeat_a.size()[1],
repeat_a.size()[2])
inp = torch.cat((repeat_s, samples, repeat_a), dim=2)
baseline = self.critic(inp).mean(0)
coma = (q_out - baseline).detach()
actor_loss -= (log_probs[0].view(coma.size()) * (coma)).mean()
# train second agent
inp = torch.cat((states, actions[0].detach(), actions[1]), dim=1)
q_out = self.critic(inp)
samples = self.monte_carlo(means[1], log_stds[1].exp())
samples = self.range * torch.tanh(samples)
repeat_a = actions[0].unsqueeze(0)
repeat_a = repeat_a.expand(samples.size()[0],
repeat_a.size()[1],
repeat_a.size()[2])
inp = torch.cat((repeat_s, repeat_a, samples), dim=2)
baseline = self.critic(inp).mean(0)
coma = (q_out - baseline).detach()
actor_loss -= (log_probs[1].view(coma.size()) * (coma)).mean()
if self.homogenous:
self.actor.optimizer.zero_grad()
actor_loss.backward()
self.actor.optimizer.step()
else:
for actor in self.actor:
actor.optimizer.zero_grad()
actor_loss.backward()
for actor in self.actor:
torch.nn.utils.clip_grad_norm_(actor.parameters(),
self.clip_grad_norm)
actor.optimizer.step()
self.totalSteps += 1
self.trained = True
#UPDATE TARGET NETWORK:
if self.totalSteps % 50 == 0:
for target_param, param in zip(self.target.parameters(),
self.critic.parameters()):
target_param.data.copy_(param.data)
return
class DoubleQ(Agent):
def __init__(self, params, name, task, load_path=None):
super(DoubleQ, self).__init__(params, name, task)
self.dual = self.vPars['dual']
if self.trainMode:
if self.dual:
self.tarNet = DualNetwork(self.vPars, self.vTrain)
self.valueNet = DualNetwork(self.vPars, self.vTrain)
else:
self.tarNet = Network(self.vPars, self.vTrain)
self.valueNet = Network(self.vPars, self.vTrain)
for target_param, param in zip(self.tarNet.parameters(),
self.valueNet.parameters()):
target_param.data.copy_(param.data)
else:
self.valueNet = Network(self.vPars, self.vTrain)
self.valueNet.load_state_dict(torch.load(load_path))
self.out_n = self.vPars['neurons'][-1]
self.replaceCounter = 0
self.valueLoss = []
self.avgLoss = 0
self.expSize = self.vTrain['buffer']
self.exp = Memory(size=self.expSize)
self.beta = self.vPars['beta']
self.priority = self.vTrain['priority']
self.priorities = []
self.alpha = .7
self.double = self.vTrain['double']
self.update_target_network = self.vTrain['update_target_network_every']
if 'noise' in self.vTrain:
self.noise = self.vTrain['noise']
else:
self.noise = 0
task.initAgent(self)
if not load_path:
while (not self.stop):
x = 1 + 1
task.postTraining()
def saveModel(self):
torch.save(
self.valueNet.state_dict(),
'/home/jimmy/Documents/Research/AN_Bridging/results/hierarchical_q_policy2.txt'
)
pass
def store(self, s, a, r, sprime, aprime, done):
self.exp.push(s, a, r, 1 - done, aprime, sprime)
if len(self.priorities) < self.expSize:
self.priorities.append(1)
else:
self.priorities = self.priorities[1:]
self.priorities.append(1)
def get_q(self, s):
if type(self.valueNet) == list:
model_index = np.random.randint(len(self.valueNet))
net = self.valueNet[model_index]
else:
net = self.valueNet
q = net(torch.FloatTensor(s))
q = q.detach()
return q
def get_action(self, s, testing_time=False, probabilistic=False):
i = np.random.random()
if i < self.explore and self.trainMode and not testing_time:
index = np.random.randint(self.out_n)
else:
q = self.get_q(s)
if probabilistic:
q = q.numpy()
q = q - np.max(q)
probs = np.exp(q * self.beta)
probs = probs / np.sum(probs)
index = np.random.choice(q.size, p=probs.ravel())
# print('probability chosen ', probs.ravel()[index])
else:
index = np.argmax(q.numpy())
self.explore = max(.2, self.explore * .9997)
return index
def get_q_and_q_tar(self, states, actions, nextStates, rewards, masks):
qValues = self.valueNet(
torch.FloatTensor(states).squeeze(1)) #pass in. Processing implied
q = torch.gather(
qValues, 1,
torch.LongTensor(actions).unsqueeze(1)) #get q values of actions
qnext = self.tarNet(torch.FloatTensor(nextStates))
qnext = qnext.squeeze(1).detach() #pass in
if self.double:
qNextDouble = self.valueNet(torch.FloatTensor(nextStates))
qNextDouble = qNextDouble.squeeze(1).detach() #pass in
qnext = torch.gather(
qnext, 1, torch.LongTensor(qNextDouble.argmax(1).unsqueeze(1)))
qtar = torch.FloatTensor(rewards).squeeze(
1) + self.discount * torch.Tensor(masks).unsqueeze(1) * qnext
else:
qtar = torch.FloatTensor(rewards) + self.discount * torch.Tensor(
masks).unsqueeze(1) * qnext.max(1)[0].view(
self.batch_size, 1) #calculate target
return q, qtar
def train(self, override=False):
if len(self.exp) >= 500 or override:
if self.priority:
loss = 0
weights = []
errors = []
assert len(self.priorities) == len(self.exp)
for i in range(self.batch_size):
probs = np.array(
[math.pow(p, self.alpha) for p in self.priorities])
probs = probs / np.sum(probs)
choice = np.random.choice(len(self.priorities),
p=probs,
size=1)
weights.append(
math.pow(
len(self.priorities) *
self.priorities[int(np.asscalar(choice))],
-self.beta))
states, actions, rewards, masks, _, nextStates, _, _, _ = self.exp.get_transitions(
choice)
q, qtar = self.get_q_and_q_tar(states, actions, nextStates,
rewards, masks)
td = qtar - q
self.priorities[int(np.asscalar(choice))] = abs(td[:, 0])
errors.append(self.valueNet.get_loss(q, qtar))
max_weight = max(weights)
weights = [w / max_weight for w in weights]
val_loss = sum([w * e for w, e in zip(weights, errors)])
else:
states, actions, rewards, masks, _, nextStates, _, _, _ = self.exp.sample(
batch=self.batch_size)
if self.replaceCounter % self.update_target_network == 0:
self.tarNet.load_state_dict(self.valueNet.state_dict())
self.replaceCounter = 0
if self.noise:
states = np.array(states)
states = states + np.random.normal(0, self.noise,
states.shape)
q, qtar = self.get_q_and_q_tar(states, actions, nextStates,
rewards, masks)
val_loss = self.valueNet.get_loss(q, qtar)
self.valueNet.optimizer.zero_grad()
val_loss.backward()
self.valueNet.optimizer.step()
self.replaceCounter += 1
self.totalSteps += 1
return val_loss
class SAC(Agent):
def __init__(self, params, name, task):
super(SAC, self).__init__(params, name, task)
self.aPars = params['actPars']
self.aTrain = params['actTrain']
self.qPars = params['qPars']
self.qTrain = params['qTrain']
if self.trainMode:
self.QNet = Network(self.qPars, self.qTrain).to(device)
self.VNet = Network(self.vPars, self.vTrain).to(device)
self.VTar = Network(self.vPars, self.vTrain).to(device)
self.policyNet = SoftNetwork(self.aPars, self.aTrain).to(device)
else:
print('Not implemented')
for target_param, param in zip(self.VTar.parameters(),
self.VNet.parameters()):
target_param.data.copy_(param)
self.expSize = self.vTrain['buffer']
self.actions = self.aPars['neurons'][-1]
self.state = self.aPars['neurons'][0]
self.exp = ReplayBuffer(self.expSize, self.actions, np.float32,
self.state, np.float32)
task.initAgent(self)
while (not self.stop):
x = 1 + 1
task.postTraining()
def load_nets(self):
pass
def saveModel(self):
pass
def get_action(self, s):
action, _, _, _, _ = self.policyNet(torch.FloatTensor(s))
action = np.ravel(action.detach().numpy())
return action
def send_to_device(self, s, a, r, next_s, d):
s = torch.FloatTensor(s).to(device)
a = torch.FloatTensor(a).to(device)
r = torch.FloatTensor(r).unsqueeze(1).to(device)
next_s = torch.FloatTensor(next_s).to(device)
d = torch.FloatTensor(np.float32(d)).unsqueeze(1).to(device)
return s, a, r, next_s, d
def train(self):
if len(self.exp) > 750:
s, a, r, next_s, d = self.exp.sample_batch(self.batch_size)
s, a, r, next_s, d = self.send_to_device(s, a, r, next_s, d)
q = self.QNet(torch.cat([s, a], dim=1))
v = self.VNet(s)
new_a, log_prob, z, mean, log_std = self.policyNet(s)
target_v = self.VTar(next_s)
next_q = r + (1 - d) * self.discount * target_v
q_loss = self.QNet.get_loss(q, next_q.detach())
new_q = self.QNet(torch.cat([s, new_a], dim=1))
next_v = new_q - log_prob * self.alpha
v_loss = self.VNet.get_loss(v, next_v.detach())
target = new_q - v
actor_loss = (log_prob *
(log_prob * self.alpha - target).detach()).mean()
mean_loss = 1e-3 * mean.pow(2).mean()
std_loss = 1e-3 * log_std.pow(2).mean()
actor_loss += mean_loss + std_loss
self.VNet.optimizer.zero_grad()
v_loss.backward()
self.VNet.optimizer.step()
self.QNet.optimizer.zero_grad()
q_loss.backward()
self.QNet.optimizer.step()
self.policyNet.optimizer.zero_grad()
actor_loss.backward()
self.policyNet.optimizer.step()
for target_param, param in zip(self.VTar.parameters(),
self.VNet.parameters()):
target_param.data.copy_(target_param.data * (1.0 - 5 * 1e-3) +
param.data * 5 * 1e-3)
self.totalSteps += 1
class Counter(object):
def __init__(self, params, name, task):
self.name = name
self.task = task
self.vTrain = params['valTrain']
self.vPars = params['valPars']
self.aTrain = params['actTrain']
self.aPars = params['actPars']
self.agents = params['agents']
self.pubs = OrderedDict()
for key in self.agents.keys():
bot = self.agents[key]
self.pubs[key] = rospy.Publisher(bot['pub'], Vector3, queue_size=1)
rospy.Subscriber("/finished", Int8, self.receiveDone, queue_size=1)
self.valueLoss = []
self.actorLoss = []
self.h_state_n = self.aPars['h_state_n']
self.x_state_n = self.aPars['x_state_n']
self.u_n = self.aPars['u_n']
self.clip_grad_norm = self.aTrain['clip']
self.homogenous = self.aPars['share_params']
self.critic = Network(self.vPars, self.vTrain).to(device)
self.target = Network(self.vPars, self.vTrain).to(device)
if self.homogenous:
self.actor = CounterActor(self.aPars, self.aTrain).to(device)
else:
self.actor = [
CounterActor(self.aPars, self.aTrain)
for i in range(len(self.agents))
]
for target_param, param in zip(self.target.parameters(),
self.critic.parameters()):
target_param.data.copy_(param.data)
self.clip_grad_norm = self.aTrain['clip']
self.trainMode = self.vPars['trainMode']
self.batch_size = self.vTrain['batch']
self.discount = self.vTrain['gamma']
self.temp_second = None
self.temp_first = None
self.td_lambda = 0 # TEST: this is because we are doing ER off-policy
self.tau = .01
self.stop = False
self.trained = False
self.exp = Memory()
self.totalSteps = 0
self.reset()
task.initAgent(self)
while (not self.stop):
x = 1 + 1
task.postTraining()
def receiveDone(self, message):
if message.data == 1: #all iterations are done. Check manager.py
self.stop = True
if message.data == 2: #timed out. Check manager.py
self.task.restartProtocol(restart=1)
def get_action(self, s_true, s_split):
if self.homogenous:
policy1 = self.actor(torch.FloatTensor(s_split[0]))
a1 = np.asscalar(self.choose(policy1))
policy2 = self.actor(torch.FloatTensor(s_split[1]))
a2 = np.asscalar(self.choose(policy2))
else:
policy1 = self.actor[0](torch.FloatTensor(s_split[0]))
a1 = self.choose(policy1)
policy2 = self.actor[1](torch.FloatTensor(s_split[1]))
a2 = self.choose(policy2)
# THIS IS A TEST
a1 = 0
#print(policy1)
#print(policy2)
#print('')
return [a1, a2]
def choose(self, policies):
m = Categorical(policies)
action = m.sample()
action = action.data.cpu().numpy()
return action
def saveModel(self):
pass
def store(self, s, a, r, sprime, aprime, done, local, next_local):
self.exp.push(s, a, r, 1 - done, aprime, sprime, local, next_local,
None)
def reset(self):
self.train(True)
if self.trained:
self.actor.eps = max(.05, self.actor.eps - .003)
self.trained = False
self.temp_first, self.temp_second = (None, None)
self.h = [
torch.zeros((1, 1, self.h_state_n))
for i in range(len(self.agents))
]
self.prevAction = [-1, -1]
return
def zipStack(self, data):
data = zip(*data)
data = [torch.stack(d).squeeze().to(device) for d in data]
return data
def get_lambda_targets(self, rewards, mask, gamma, target_qs):
target_qs = target_qs.squeeze()
ret = target_qs.new_zeros(*target_qs.shape)
ret[-1] = rewards[-1] + target_qs[-1] * mask[-1]
for t in range(ret.size()[0] - 2, -1, -1):
ret[t] = mask[t] * (self.td_lambda * gamma * ret[t + 1]) + (
rewards[t] +
(1 - self.td_lambda) * gamma * target_qs[t] * mask[t])
return ret.unsqueeze(1)
def train(self, episode_done=False):
if len(self.exp) > self.batch_size:
transition = self.exp.sample(self.batch_size)
states = torch.squeeze(torch.Tensor(transition.state)).to(device)
states_next = torch.squeeze(torch.Tensor(
transition.next_state)).to(device)
actions = torch.Tensor(transition.action).float().to(device)
rewards = torch.Tensor(transition.reward).to(device)
masks = torch.Tensor(transition.mask).to(device)
local = self.zipStack(transition.local)
next_local = self.zipStack(transition.next_local)
actions_next = []
for s in next_local:
next_policy = self.actor(s)
next_action = self.choose(next_policy)
actions_next.append(torch.Tensor(next_action))
'''# Critic Update
ID = torch.Tensor(states.size()[0], 1).fill_(-1)
inp = torch.cat((states_next, actions_next[1].unsqueeze(1), ID), dim = 1)
q_tar = self.target(inp).detach().gather(1, actions_next[0].long().unsqueeze(1))
q_tar = self.get_lambda_targets(rewards.squeeze(), masks.squeeze(), self.discount, q_tar)
inp = torch.cat((states, actions[:, 1].unsqueeze(1), ID), dim = 1)
q = self.critic(inp)
q = q.gather(1, actions[:, 0].long().unsqueeze(1))
loss = self.critic.get_loss(q, q_tar)
self.critic.optimizer.zero_grad()
loss.backward()
self.critic.optimizer.step()'''
ID = torch.Tensor(states.size()[0], 1).fill_(1)
inp = torch.cat((states_next, actions_next[0].unsqueeze(1), ID),
dim=1)
q_tar = self.target(inp).detach().gather(
1, actions_next[1].long().unsqueeze(1)) # .max(1)?
q_tar = self.get_lambda_targets(rewards.squeeze(), masks.squeeze(),
self.discount, q_tar)
inp = torch.cat((states, actions[:, 0].unsqueeze(1), ID), dim=1)
q = self.critic(inp)
q = q.gather(1, actions[:, 1].long().unsqueeze(1))
loss = self.critic.get_loss(q, q_tar)
self.critic.optimizer.zero_grad()
loss.backward()
self.critic.optimizer.step()
actor_loss = 0
# Actor Update. Consider doing new_actions
policies = []
new_actions = []
for s in local:
policy = self.actor(s)
policies.append(policy)
new_action = self.choose(policy)
new_actions.append(torch.Tensor(new_action))
'''ID = torch.Tensor(states.size()[0], 1).fill_(-1)
inp = torch.cat((states, new_actions[1].unsqueeze(1), ID), dim = 1)
q_out = self.critic(inp) #batch x num_actions
policy = policies[0] #batch x num_actions
mult = q_out * policy
baseline = torch.sum(mult, 1).unsqueeze(1)
q_taken = q_out.gather(1, new_actions[0].long().unsqueeze(1))
coma = (q_taken - baseline).detach()
probs_taken = policy.gather(1, new_actions[0].long().unsqueeze(1))
loss = -(torch.log(probs_taken) * coma).mean()
actor_loss += loss '''
ID = torch.Tensor(states.size()[0], 1).fill_(1)
inp = torch.cat((states, new_actions[0].unsqueeze(1), ID), dim=1)
q_out = self.critic(inp) #batch x num_actions
policy = policies[1] #batch x num_actions
mult = q_out * policy
baseline = torch.sum(mult, 1).unsqueeze(1)
q_taken = q_out.gather(1, new_actions[1].long().unsqueeze(1))
coma = (q_taken - baseline).detach()
probs_taken = policy.gather(1, new_actions[1].long().unsqueeze(1))
loss = -(torch.log(probs_taken) * coma).mean()
actor_loss += loss
self.actorLoss.append(actor_loss)
if self.homogenous:
self.actor.optimizer.zero_grad()
actor_loss.backward()
nn.utils.clip_grad_norm_(self.actor.parameters(), 1)
self.actor.optimizer.step()
else:
for actor in self.actor:
actor.optimizer.zero_grad()
actor_loss.backward()
for actor in self.actor:
actor.optimizer.step()
self.totalSteps += 1
# self.exp = Memory()
self.trained = True
#UPDATE TARGET NETWORK:
if self.totalSteps % 1 == 0: # THIS IS A TEST
for target_param, param in zip(self.target.parameters(),
self.critic.parameters()):
target_param.data.copy_((1 - self.tau) * target_param +
self.tau * param.data)
return
class DoubleQ(Agent):
def __init__(self, params, name, task, load_path=None):
super(DoubleQ, self).__init__(params, name, task)
self.dual = self.vPars['dual']
if self.trainMode:
if self.dual:
self.tarNet = DualNetwork(self.vPars, self.vTrain)
self.valueNet = DualNetwork(self.vPars, self.vTrain)
else:
self.tarNet = Network(self.vPars, self.vTrain)
self.valueNet = Network(self.vPars, self.vTrain)
for target_param, param in zip(self.tarNet.parameters(),
self.valueNet.parameters()):
target_param.data.copy_(param.data)
else:
self.valueNet = Network(self.vPars, self.vTrain)
paths = [
'/home/austinnguyen517/Documents/Research/BML/MultiRobot/AN_Bridging/push_in_hole.txt',
'/home/austinnguyen517/Documents/Research/BML/MultiRobot/AN_Bridging/push_in_hole2.txt'
]
if not load_path:
self.valueNet = []
for path in paths:
self.valueNet.append(Network(self.vPars, self.vTrain))
self.valueNet[-1].load_state_dict(torch.load(path))
else:
self.valueNet.load_state_dict(torch.load(load_path))
self.out_n = self.vPars['neurons'][-1]
self.replaceCounter = 0
self.valueLoss = []
self.avgLoss = 0
self.expSize = self.vTrain['buffer']
self.exp = Memory(size=self.expSize)
self.double = self.vTrain['double']
task.initAgent(self)
if not load_path:
while (not self.stop):
x = 1 + 1
task.postTraining()
def saveModel(self):
torch.save(
self.valueNet.state_dict(),
'/home/austinnguyen517/Documents/Research/BML/MultiRobot/AN_Bridging/box_push_hierarchical_q_policy.txt'
)
pass
def store(self, s, a, r, sprime, aprime, done):
self.exp.push(s, a, r, 1 - done, aprime, sprime)
def get_action(self, s):
i = np.random.random()
if i < self.explore and self.trainMode:
index = np.random.randint(self.out_n)
else:
if type(self.valueNet) == list:
model_index = np.random.randint(len(self.valueNet))
net = self.valueNet[model_index]
else:
net = self.valueNet
q = net(torch.FloatTensor(s))
#print(q)
q = q.detach()
index = np.argmax(q.numpy())
self.explore = max(.1, self.explore * .9997)
return index
def train(self):
if len(self.exp) >= 500:
states, actions, rewards, masks, _, nextStates, _, _, _ = self.exp.sample(
batch=self.batch_size)
if self.replaceCounter % 500 == 0: # THIS IS SET TO 200 FOR BOX PUSH TASK...SLOPE IS 500
self.tarNet.load_state_dict(self.valueNet.state_dict())
self.replaceCounter = 0
qValues = self.valueNet(torch.FloatTensor(states).squeeze(
1)) #pass in. Processing implied
q = torch.gather(qValues, 1,
torch.LongTensor(actions).unsqueeze(
1)) #get q values of actions
qnext = self.tarNet(torch.FloatTensor(nextStates))
qnext = qnext.squeeze(1).detach() #pass in
if self.double:
qNextDouble = self.valueNet(torch.FloatTensor(nextStates))
qNextDouble = qNextDouble.squeeze(1).detach() #pass in
qnext = torch.gather(
qnext, 1,
torch.LongTensor(qNextDouble.argmax(1).unsqueeze(1)))
qtar = torch.FloatTensor(rewards).squeeze(
1
) + self.discount * torch.Tensor(masks).unsqueeze(1) * qnext
else:
qtar = torch.FloatTensor(
rewards) + self.discount * torch.Tensor(
masks).unsqueeze(1) * qnext.max(1)[0].view(
self.batch_size, 1) #calculate target
val_loss = self.valueNet.get_loss(q, qtar)
self.valueNet.optimizer.zero_grad()
val_loss.backward()
self.valueNet.optimizer.step()
self.replaceCounter += 1
self.totalSteps += 1