def main():
# setup logger
if args.resume_dir == "":
date = str(datetime.datetime.now())
date = date[:date.rfind(":")].replace("-", "") \
.replace(":", "") \
.replace(" ", "_")
log_dir = os.path.join(args.log_root, "log_" + date)
else:
log_dir = args.resume_dir
hparams_file = os.path.join(log_dir, "hparams.json")
checkpoints_dir = os.path.join(log_dir, "checkpoints")
if not os.path.exists(log_dir):
os.makedirs(log_dir)
if not os.path.exists(checkpoints_dir):
os.makedirs(checkpoints_dir)
if args.resume_dir == "":
# write hparams
with open(hparams_file, "w") as f:
json.dump(args.__dict__, f, indent=2)
log_file = os.path.join(log_dir, "log_train.txt")
logger = Logger(log_file)
# logger.info(args)
logger.info("The args corresponding to training process are: ")
for (key, value) in vars(args).items():
logger.info("{key:20}: {value:}".format(key=key, value=value))
actor_critic = ActorCritic(args, log_dir, checkpoints_dir)
actor_critic.train()
def train(rank,
args,
shared_model,
counter,
lock,
optimizer=None,
select_sample=True):
torch.manual_seed(args.seed + rank)
print("Process No : {} | Sampling : {}".format(rank, select_sample))
FloatTensor = torch.cuda.FloatTensor if args.use_cuda else torch.FloatTensor
DoubleTensor = torch.cuda.DoubleTensor if args.use_cuda else torch.DoubleTensor
ByteTensor = torch.cuda.ByteTensor if args.use_cuda else torch.ByteTensor
env = create_mario_env(args.env_name)
env.seed(args.seed + rank)
model = ActorCritic(env.observation_space.shape[0], len(ACTIONS))
if args.use_cuda:
model.cuda()
if optimizer is None:
optimizer = optim.Adam(shared_model.parameters(), lr=args.lr)
model.train()
state = env.reset()
state = torch.from_numpy(state)
done = True
episode_length = 0
for num_iter in count():
if rank == 0:
env.render()
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_path)
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_path)
# Sync with the shared model
model.load_state_dict(shared_model.state_dict())
if done:
cx = Variable(torch.zeros(1, 512)).type(FloatTensor)
hx = Variable(torch.zeros(1, 512)).type(FloatTensor)
else:
cx = Variable(cx.data).type(FloatTensor)
hx = Variable(hx.data).type(FloatTensor)
values = []
log_probs = []
rewards = []
entropies = []
reason = ''
for step in range(args.num_steps):
episode_length += 1
state_inp = Variable(state.unsqueeze(0)).type(FloatTensor)
value, logit, (hx, cx) = model((state_inp, (hx, cx)))
prob = F.softmax(logit, dim=-1)
log_prob = F.log_softmax(logit, dim=-1)
entropy = -(log_prob * prob).sum(-1, keepdim=True)
entropies.append(entropy)
if select_sample:
action = prob.multinomial().data
else:
action = prob.max(-1, keepdim=True)[1].data
log_prob = log_prob.gather(-1, Variable(action))
action_out = ACTIONS[action][0, 0]
# print("Process: {} Action: {}".format(rank, str(action_out)))
state, reward, done, _ = env.step(action_out)
done = done or episode_length >= args.max_episode_length
reward = max(min(reward, 50), -50)
with lock:
counter.value += 1
if done:
episode_length = 0
env.change_level(0)
state = env.reset()
print("Process {} has completed.".format(rank))
env.locked_levels = [False] + [True] * 31
state = torch.from_numpy(state)
values.append(value)
log_probs.append(log_prob)
rewards.append(reward)
if done:
break
R = torch.zeros(1, 1)
if not done:
state_inp = Variable(state.unsqueeze(0)).type(FloatTensor)
value, _, _ = model((state_inp, (hx, cx)))
R = value.data
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)
# Generalized Advantage Estimataion
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) - args.entropy_coef * entropies[i]
total_loss = policy_loss + args.value_loss_coef * value_loss
print("Process {} loss :".format(rank), total_loss.data)
# print("Process: {} Episode: {}".format(rank, str(episode_length)))
optimizer.zero_grad()
(total_loss).backward()
torch.nn.utils.clip_grad_norm(model.parameters(), args.max_grad_norm)
ensure_shared_grads(model, shared_model)
optimizer.step()
print("Process {} closed.".format(rank))
def train(rank, args, shared_model, shared_scme, counter, lock, optimizer=None, select_sample=True):
torch.manual_seed(args.seed + rank)
print("Process No : {} | Sampling : {}".format(rank, select_sample))
FloatTensor = torch.FloatTensor# torch.cuda.FloatTensor if args.use_cuda else torch.FloatTensor
DoubleTensor = torch.DoubleTensor# torch.cuda.DoubleTensor if args.use_cuda else torch.DoubleTensor
ByteTensor = torch.ByteTensor# torch.cuda.ByteTensor if args.use_cuda else torch.ByteTensor
savefile = os.getcwd() + '/save/scmemi_'+ args.reward_type +'/train_reward.csv'
saveweights = os.getcwd() + '/save/scmemi_'+ args.reward_type +'/mario_a3c_params.pkl'
env = create_mario_env(args.env_name, args.reward_type)
#env.seed(args.seed + rank)
model = ActorCritic(env.observation_space.shape[0], len(ACTIONS))
if optimizer is None:
optimizer = optim.Adam(list(shared_model.parameters()) + list(shared_scme.parameters()), lr=args.lr)
scme_model = SCME(env.observation_space.shape[0], len(ACTIONS))
model.train()
scme_model.train()
state = env.reset()
cum_rew = 0
state = torch.from_numpy(state)
done = True
episode_length = 0
for num_iter in count():
#env.render()
if rank == 0:
if num_iter % args.save_interval == 0 and num_iter > 0:
print ("Saving model at :" + saveweights)
torch.save(shared_model.state_dict(), saveweights)
torch.save(shared_scme.state_dict(), saveweights[:-4] + '_scme.pkl')
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 :" + saveweights)
torch.save(shared_model.state_dict(), saveweights)
torch.save(shared_scme.state_dict(), saveweights[:-4] + '_scme.pkl')
# Sync with the shared model
model.load_state_dict(shared_model.state_dict())
scme_model.load_state_dict(shared_scme.state_dict())
if done:
cx = Variable(torch.zeros(1, 512)).type(FloatTensor)
hx = Variable(torch.zeros(1, 512)).type(FloatTensor)
else:
cx = Variable(cx.data).type(FloatTensor)
hx = Variable(hx.data).type(FloatTensor)
values = []
log_probs = []
rewards = []
entropies = []
vae_losses = []
cur_losses = []
mi_losses = []
#reason =''
for step in range(args.num_steps):
episode_length += 1
state_inp = Variable(state.unsqueeze(0)).type(FloatTensor)
value, logit, (hx, cx) = model((state_inp, (hx, cx)))
prob = F.softmax(logit, dim=-1)
log_prob = F.log_softmax(logit, dim=-1)
entropy = -(log_prob * prob).sum(-1, keepdim=True)
entropies.append(entropy)
if select_sample:
action = prob.multinomial(1).data
else:
action = prob.max(-1, keepdim=True)[1].data
log_prob = log_prob.gather(-1, Variable(action))
action_out = int(action[0, 0].data.numpy())
state, reward, done, info = env.step(action_out)
cum_rew = cum_rew + reward
action_one_hot = (torch.eye(len(ACTIONS))[action_out]).view(1,-1)
next_state_inp = Variable(torch.from_numpy(state).unsqueeze(0)).type(FloatTensor)
pred_z, mi, mi1, actual_z, xt1_hat, xt1, xt1_mu, xt1_logvar = scme_model((state_inp, next_state_inp, action_one_hot))
vae_loss = loss_function(xt1_hat, xt1, xt1_mu, xt1_logvar)
cur_loss = ((pred_z - actual_z).pow(2)).sum(-1, keepdim=True)/2/50
mi_loss = mutual(mi, mi1).sum(-1, keepdim=True)/10
done = done or episode_length >= args.max_episode_length
cur_reward = (args.alpha*cur_loss).data.numpy()[0,0]
mi_reward = (args.beta*mi_loss).data.numpy()
reward = cur_reward + reward + mi_reward
reward = max(min(reward, 50), -5)
with lock:
counter.value += 1
if done:
episode_length = 0
# env.change_level(0)
state = env.reset()
with open(savefile[:-4]+'_{}.csv'.format(rank), 'a', newline='') as sfile:
writer = csv.writer(sfile)
writer.writerows([[cum_rew, info['x_pos']/x_norm]])
cum_rew = 0
# print ("Process {} has completed.".format(rank))
# env.locked_levels = [False] + [True] * 31
state = torch.from_numpy(state)
values.append(value)
log_probs.append(log_prob)
rewards.append(reward)
vae_losses.append(vae_loss)
cur_losses.append(cur_loss)
mi_losses.append(mi_loss)
if done:
break
R = torch.zeros(1, 1)
if not done:
state_inp = Variable(state.unsqueeze(0)).type(FloatTensor)
value, _, _ = model((state_inp, (hx, cx)))
R = value.data
values.append(Variable(R).type(FloatTensor))
policy_loss = 0
value_loss = 0
scme_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)
# Generalized Advantage Estimataion
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) - args.entropy_coef * entropies[i]
scme_loss = 0.01*vae_losses[i] + cur_losses[i] - mi_losses[i]
total_loss = args.lambd*(policy_loss + args.value_loss_coef * value_loss)
# print ("Process {} loss :".format(rank), total_loss.data)
optimizer.zero_grad()
# cur_optimizer.zero_grad()
(total_loss + scme_loss).backward()
# (curiosity_loss).backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
torch.nn.utils.clip_grad_norm_(scme_model.parameters(), args.max_grad_norm)
ensure_shared_grads(model, shared_model)
ensure_shared_grads(scme_model, shared_scme)
optimizer.step()
