def train(rank, args, shared_model, counter, lock, optimizer=None):
FloatTensor = torch.cuda.FloatTensor if args.use_cuda else torch.FloatTensor
env = gym.make("FetchPickAndPlace-v1")
env2 = gym.wrappers.FlattenDictWrapper(env, dict_keys=['observation', 'desired_goal'])
model = Actor()
model2 = second()
if args.use_cuda:
model.cuda()
model2.cuda()
if os.path.isfile(args.save_path2):
print('Loading second parametets ...')
pretrained_dict = torch.load(args.save_path2)
model_dict2 = model2.state_dict()
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict2}
model_dict2.update(pretrained_dict)
model2.load_state_dict(model_dict2)
for p in model.fc1.parameters():
p.requires_grad = False
for p in model.fc2.parameters():
p.requires_grad = False
if optimizer is None:
optimizer = optim.Adam(shared_model.parameters(), lr=args.lr)
model.train()
model2.eval()
done = True
for num_iter in count():
with lock:
counter.value += 1
#print(num_iter, counter.value)
lastObs = env.reset()
goal = lastObs['desired_goal']
objectPos = lastObs['observation'][3:6]
object_rel_pos = lastObs['observation'][6:9]
object_oriented_goal = object_rel_pos.copy()
object_oriented_goal[2] += 0.03 # first make the gripper go slightly above the object
timeStep = 0 #count the total number of timesteps
if rank == 0:
if num_iter % args.save_interval == 0 and num_iter > 0:
#print ("Saving model at :" + args.save_path)
torch.save(shared_model.state_dict(), args.save_path1)
if num_iter % (args.save_interval * 2.5) == 0 and num_iter > 0 and rank == 1: # Second saver in-case first processes crashes
#print ("Saving model for process 1 at :" + args.save_path)
torch.save(shared_model.state_dict(), args.save_path1)
model.load_state_dict(shared_model.state_dict())
values, log_probs, rewards, entropies = [], [], [], []
if done:
cx = Variable(torch.zeros(1, 32)).type(FloatTensor)
hx = Variable(torch.zeros(1, 32)).type(FloatTensor)
else:
cx = Variable(cx.data).type(FloatTensor)
hx = Variable(hx.data).type(FloatTensor)
state_inp = torch.from_numpy(env2.observation(lastObs)).type(FloatTensor)
#criterion = nn.MSELoss()
value, y, (hx, cx) = model(state_inp, hx, cx)
prob = F.softmax(y)
log_prob = F.log_softmax(y, dim=-1)
act_model = prob.max(-1, keepdim=True)[1].data
entropy = -(log_prob * prob).sum(-1, keepdim=True)
log_prob = log_prob.gather(-1, Variable(act_model))
action_out = act_model.to(torch.device("cpu"))
#action_out = torch.tensor([[1]])
entropies.append(entropy), log_probs.append(log_prob), values.append(value)
#print(action_out)
while np.linalg.norm(object_oriented_goal) >= 0.015 and timeStep <= env._max_episode_steps:
#env.render()
action = [0, 0, 0, 0, 0, 0]
act_tensor= act(state_inp, action_out, model2)
#print(act_tensor)
for i in range(len(object_oriented_goal)):
action[i] = act_tensor[i].cpu().detach().numpy()
object_oriented_goal = object_rel_pos.copy()
object_oriented_goal[2] += 0.03
action[3] = 0.05
obsDataNew, reward, done, info = env.step(action)
timeStep += 1
objectPos = obsDataNew['observation'][3:6]
object_rel_pos = obsDataNew['observation'][6:9]
state_inp = torch.from_numpy(env2.observation(obsDataNew)).type(FloatTensor)
if timeStep >= env._max_episode_steps:
reward = torch.Tensor([-1.0]).type(FloatTensor)
break
if timeStep < env._max_episode_steps:
reward = torch.Tensor([1.0]).type(FloatTensor)
rewards.append(reward)
value, y, (hx, cx) = model(state_inp, hx, cx)
prob = F.softmax(y)
log_prob = F.log_softmax(y, dim=-1)
act_model = prob.max(-1, keepdim=True)[1].data
entropy = -(log_prob * prob).sum(-1, keepdim=True)
log_prob = log_prob.gather(-1, Variable(act_model))
action_out = act_model.to(torch.device("cpu"))
entropies.append(entropy), log_probs.append(log_prob), values.append(value)
#action_out = torch.tensor([[0]])
while np.linalg.norm(object_rel_pos) >= 0.005 and timeStep <= env._max_episode_steps :
#env.render()
action = [0, 0, 0, 0, 0, 0]
act_tensor= act(state_inp, action_out, model2)
for i in range(len(object_oriented_goal)):
action[i] = act_tensor[i].cpu().detach().numpy()
action[3]= -0.01
if action_out == 0:
action[4] = act_tensor[3].cpu().detach().numpy()
obsDataNew, reward, done, info = env.step(action)
timeStep += 1
objectPos = obsDataNew['observation'][3:6]
object_rel_pos = obsDataNew['observation'][6:9]
state_inp = torch.from_numpy(env2.observation(obsDataNew)).type(FloatTensor)
if timeStep >= env._max_episode_steps:
reward = torch.Tensor([-1.0]).type(FloatTensor)
break
if timeStep < env._max_episode_steps:
reward = torch.Tensor([1.0]).type(FloatTensor)
rewards.append(reward)
value, y, (hx, cx) = model(state_inp, hx, cx)
prob = F.softmax(y)
log_prob = F.log_softmax(y, dim=-1)
act_model = prob.max(-1, keepdim=True)[1].data
entropy = -(log_prob * prob).sum(-1, keepdim=True)
log_prob = log_prob.gather(-1, Variable(act_model))
action_out = act_model.to(torch.device("cpu"))
entropies.append(entropy), log_probs.append(log_prob), values.append(value)
#action_out = torch.tensor([[2]])
while np.linalg.norm(goal - objectPos) >= 0.01 and timeStep <= env._max_episode_steps :
#env.render()
action = [0, 0, 0, 0, 0, 0]
act_tensor= act(state_inp, action_out, model2)
for i in range(len(goal - objectPos)):
action[i] = act_tensor[i].cpu().detach().numpy()
action[3] = -0.01
obsDataNew, reward, done, info = env.step(action)
timeStep += 1
state_inp = torch.from_numpy(env2.observation(obsDataNew)).type(FloatTensor)
objectPos = obsDataNew['observation'][3:6]
object_rel_pos = obsDataNew['observation'][6:9]
if timeStep >= env._max_episode_steps:
break
while True: #limit the number of timesteps in the episode to a fixed duration
#env.render()
action = [0, 0, 0, 0, 0, 0]
action[3] = -0.01 # keep the gripper closed
obsDataNew, reward, done, info = env.step(action)
timeStep += 1
objectPos = obsDataNew['observation'][3:6]
object_rel_pos = obsDataNew['observation'][6:9]
if timeStep >= env._max_episode_steps: break
if info['is_success'] == 1.0:
reward = torch.Tensor([1.0]).type(FloatTensor)
else:
reward = torch.Tensor([-1.0]).type(FloatTensor)
rewards.append(reward)
R = torch.zeros(1, 1)
values.append(Variable(R).type(FloatTensor))
policy_loss = 0
value_loss = 0
R = Variable(R).type(FloatTensor)
gae = torch.zeros(1, 1).type(FloatTensor)
for i in reversed(range(len(rewards))):
R = args.gamma * R + rewards[i]
advantage = R - values[i]
value_loss = value_loss + 0.5 * advantage.pow(2)
delta_t = rewards[i] + args.gamma * \
values[i + 1].data - values[i].data
gae = gae * args.gamma * args.tau + delta_t
policy_loss = policy_loss - \
log_probs[i] * Variable(gae).type(FloatTensor)
total_loss = policy_loss + args.value_loss_coef * value_loss
optimizer.zero_grad()
(total_loss).backward(retain_graph=True)
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
ensure_shared_grads(model, shared_model)
optimizer.step()
type=float,
default=0.9,
help='discount factor for rewards (default: 0.9)')
parser.add_argument('--tau',
type=float,
default=1.00,
help='parameter for GAE (default: 1.00)')
args = parser.parse_args()
model = Actor()
model2 = second()
if args.use_cuda:
model.cuda()
model2.cuda()
torch.cuda.manual_seed_all(21)
optimizer = optim.Adam(model.parameters(), lr=0.0001)
if os.path.isfile(args.save_path1):
print('Loading A3C parametets ...')
model.load_state_dict(torch.load(args.save_path1))
if os.path.isfile(args.save_path2):
print('Loading second parametets ...')
pretrained_dict = torch.load(args.save_path2)
model_dict2 = model2.state_dict()
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict2
}
model_dict2.update(pretrained_dict)
model2.load_state_dict(model_dict2)
class DDPG:
def __init__(self, env):
self.env = env
self.stateDim = obs2state(env.reset().observation).size()[1]
self.actionDim = env.action_spec().shape[0]
self.actor = Actor(self.env).cuda()
self.critic = Critic(self.env).cuda()
self.targetActor = deepcopy(Actor(self.env)).cuda()
self.targetCritic = deepcopy(Critic(self.env)).cuda()
self.actorOptim = optim.Adam(self.actor.parameters(), lr=ACTOR_LR)
self.criticOptim = optim.Adam(self.critic.parameters(), lr=CRITIC_LR)
self.criticLoss = nn.MSELoss()
self.noise = OUNoise(mu=np.zeros(self.actionDim), sigma=SIGMA)
self.replayBuffer = Buffer(BUFFER_SIZE)
self.batchSize = MINIBATCH_SIZE
self.checkpoint_dir = CHECKPOINT_DIR
self.discount = DISCOUNT
self.warmup = WARMUP
self.epsilon = EPSILON
self.epsilon_decay = EPSILON_DECAY
self.rewardgraph = []
self.start = 0
self.end = NUM_EPISODES
def getQTarget(self, nextStateBatch, rewardBatch, terminalBatch):
"""Inputs: Batch of next states, rewards and terminal flags of size self.batchSize
Calculates the target Q-value from reward and bootstraped Q-value of next state
using the target actor and target critic
Outputs: Batch of Q-value targets"""
targetBatch = torch.FloatTensor(rewardBatch).cuda()
nonFinalMask = torch.ByteTensor(
tuple(map(lambda s: s != True, terminalBatch)))
nextStateBatch = torch.cat(nextStateBatch)
nextActionBatch = self.targetActor(nextStateBatch)
nextActionBatch.volatile = True
qNext = self.targetCritic(nextStateBatch, nextActionBatch)
nonFinalMask = self.discount * nonFinalMask.type(
torch.cuda.FloatTensor)
targetBatch += nonFinalMask * qNext.squeeze().data
return Variable(targetBatch, volatile=False)
def updateTargets(self, target, original):
"""Weighted average update of the target network and original network
Inputs: target actor(critic) and original actor(critic)"""
for targetParam, orgParam in zip(target.parameters(),
original.parameters()):
targetParam.data.copy_((1 - TAU)*targetParam.data + \
TAU*orgParam.data)
def getMaxAction(self, curState):
"""Inputs: Current state of the episode
Returns the action which maximizes the Q-value of the current state-action pair"""
spec = self.env.action_spec()
minAct = Variable(torch.cuda.FloatTensor(spec.minimum),
requires_grad=False)
maxAct = Variable(torch.cuda.FloatTensor(spec.maximum),
requires_grad=False)
noise = self.epsilon * Variable(torch.FloatTensor(self.noise()),
volatile=True).cuda()
action = self.actor(curState)
actionNoise = action + noise
return actionNoise
def train(self):
if not os.path.exists(self.checkpoint_dir):
os.makedirs(self.checkpoint_dir)
print('Training started...')
for i in range(self.start, self.end):
time_step = self.env.reset()
ep_reward = 0
while not time_step.last():
#Visualize Training
display.clear_output(wait=True)
plt.imshow(self.env.physics.render())
plt.show()
# Get maximizing action
curState = Variable(obs2state(time_step.observation),
volatile=True).cuda()
self.actor.eval()
action = self.getMaxAction(curState)
curState.volatile = False
action.volatile = False
self.actor.train()
# Step episode
time_step = self.env.step(action.data)
nextState = Variable(obs2state(time_step.observation),
volatile=True).cuda()
reward = time_step.reward
ep_reward += reward
terminal = time_step.last()
# Update replay bufer
self.replayBuffer.append(
(curState, action, nextState, reward, terminal))
# Training loop
if len(self.replayBuffer) >= self.warmup:
curStateBatch, actionBatch, nextStateBatch, \
rewardBatch, terminalBatch = self.replayBuffer.sample_batch(self.batchSize)
curStateBatch = torch.cat(curStateBatch)
actionBatch = torch.cat(actionBatch)
qPredBatch = self.critic(curStateBatch, actionBatch)
qTargetBatch = self.getQTarget(nextStateBatch, rewardBatch,
terminalBatch)
# Critic update
self.criticOptim.zero_grad()
criticLoss = self.criticLoss(qPredBatch, qTargetBatch)
print('Critic Loss: {}'.format(criticLoss))
criticLoss.backward()
self.criticOptim.step()
# Actor update
self.actorOptim.zero_grad()
actorLoss = -torch.mean(
self.critic(curStateBatch, self.actor(curStateBatch)))
print('Actor Loss: {}'.format(actorLoss))
actorLoss.backward()
self.actorOptim.step()
# Update Targets
self.updateTargets(self.targetActor, self.actor)
self.updateTargets(self.targetCritic, self.critic)
self.epsilon -= self.epsilon_decay
if i % 20 == 0:
self.save_checkpoint(i)
self.rewardgraph.append(ep_reward)
def save_checkpoint(self, episode_num):
checkpointName = self.checkpoint_dir + 'ep{}.pth.tar'.format(
episode_num)
checkpoint = {
'episode': episode_num,
'actor': self.actor.state_dict(),
'critic': self.critic.state_dict(),
'targetActor': self.targetActor.state_dict(),
'targetCritic': self.targetCritic.state_dict(),
'actorOpt': self.actorOptim.state_dict(),
'criticOpt': self.criticOptim.state_dict(),
'replayBuffer': self.replayBuffer,
'rewardgraph': self.rewardgraph,
'epsilon': self.epsilon
}
torch.save(checkpoint, checkpointName)
def loadCheckpoint(self, checkpointName):
if os.path.isfile(checkpointName):
print("Loading checkpoint...")
checkpoint = torch.load(checkpointName)
self.start = checkpoint['episode'] + 1
self.actor.load_state_dict(checkpoint['actor'])
self.critic.load_state_dict(checkpoint['critic'])
self.targetActor.load_state_dict(checkpoint['targetActor'])
self.targetCritic.load_state_dict(checkpoint['targetCritic'])
self.actorOptim.load_state_dict(checkpoint['actorOpt'])
self.criticOptim.load_state_dict(checkpoint['criticOpt'])
self.replayBuffer = checkpoint['replayBuffer']
self.rewardgraph = checkpoint['rewardgraph']
self.epsilon = checkpoint['epsilon']
print('Checkpoint loaded')
else:
raise OSError('Checkpoint not found')
torch.cuda.manual_seed_all(30)
shared_model.share_memory()
if os.path.isfile(args.save_path1):
print('Loading A3C parametets ...')
pretrained_dict = torch.load(args.save_path1)
model_dict = shared_model.state_dict()
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict
}
model_dict.update(pretrained_dict)
shared_model.load_state_dict(model_dict)
optimizer = SharedAdam(shared_model.parameters(), lr=args.lr)
optimizer.share_memory()
for p in shared_model.fc1.parameters():
p.requires_grad = False
for p in shared_model.fc2.parameters():
p.requires_grad = False
print("No of available cores : {}".format(mp.cpu_count()))
processes = []
counter = mp.Value('i', 0)
lock = mp.Lock()
print(counter)
p = mp.Process(target=test,
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size,
random_seed).to(device)
self.actor_target = Actor(state_size, action_size,
random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target = Critic(state_size, action_size,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def step(self, state, action, reward, next_state, done, times):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
self.memory.add(state, action, reward, next_state, done)
# Learn, if enough samples are available in memory
if len(self.memory) > BATCH_SIZE and times % LEARN_EVERY == 0:
for _ in range(LEARN_TIMES):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
if add_noise:
action += self.noise.sample() * self.epsilon
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def learn(self, experiences, gamma):
"""Update policy and value parameters using given batch of experience tuples.
Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
where:
actor_target(state) -> action
critic_target(state, action) -> Q-value
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
actions_next = self.actor_target(next_states.to(device))
Q_targets_next = self.critic_target(next_states.to(device),
actions_next.to(device))
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(
Q_expected, Q_targets
) # Critic loss uses TD method for DQN (first the MSE loss then backpropagation)
# Minimize the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
clip_grad_norm_(self.critic_local.parameters(),
1) # Clip the gradient when update critic network
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(states, actions_pred).mean(
) # Actor loss uses -Q from local critic network
# Minimize the loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(self.critic_local, self.critic_target, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
# ----------------------- update epsilon and noise ----------------------- #
self.epsilon *= EPSILON_DECAY
self.noise.reset()
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model: PyTorch model (weights will be copied from)
target_model: PyTorch model (weights will be copied to)
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)