class AgentA3C(Agent):
'''Agent which leverages Actor Critic Learning'''
def __init__(self,
model_path,
dtype,
seed=451):
self._seed = seed
self._idx = 0
self._dtype = dtype
self.env = LoveLetterEnv(AgentRandom(seed), seed)
state = self.env.reset()
self._model = ActorCritic(
state.shape[0], self.env.action_space).type(dtype)
self._model.load_state_dict(torch.load(model_path))
def _move(self, game):
'''Return move which ends in score hole'''
assert game.active()
self._idx += 1
state = self.env.force(game)
state = torch.from_numpy(state).type(self._dtype)
cx = Variable(torch.zeros(1, 256).type(self._dtype), volatile=True)
hx = Variable(torch.zeros(1, 256).type(self._dtype), volatile=True)
_, logit, (hx, cx) = self._model(
(Variable(state.unsqueeze(0), volatile=True), (hx, cx)))
prob = F.softmax(logit)
action_idx = prob.max(1)[1].data.cpu().numpy()[0, 0]
player_action = self.env.action_from_index(action_idx, game)
if player_action is None:
# print("ouch")
options = Agent.valid_actions(game, self._seed + self._idx)
if len(options) < 1:
raise Exception("Unable to play without actions")
random.seed(self._seed + self._idx)
return random.choice(options)
# print("playing ", self._idx, player_action)
return player_action
class TFAgent(Agent):
'''Agent which leverages tensorflow'''
def __init__(self, model_path, seed=451):
self._seed = seed
self._idx = 0
self.env = LoveLetterEnv(AgentRandom(seed), seed)
self.vec_env = DummyVecEnv([
lambda: self.env
]) # The algorithms require a vectorized environment to run
state = self.env.reset()
self._model = PPO2(MlpPolicy,
self.vec_env,
verbose=0,
tensorboard_log="./tensorboard/")
self._model.load(model_path)
def _move(self, game):
'''Return move which ends in score hole'''
assert game.active()
self._idx += 1
state = self.env.force(game)
action_idx = self._model.predict(state, deterministic=True)[0]
player_action = self.env.action_from_index(action_idx, game)
if player_action is None:
# print("ouch")
options = Agent.valid_actions(game, self._seed + self._idx)
if len(options) < 1:
raise Exception("Unable to play without actions")
random.seed(self._seed + self._idx)
return random.choice(options)
# print("playing ", self._idx, player_action)
return player_action
parser.add_argument(
'--load-name',
default=None,
metavar='SN',
help='path/prefix for the filename to load shared model\'s parameters')
if __name__ == '__main__':
args = parser.parse_args()
torch.manual_seed(args.seed)
dtype = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.FloatTensor
env = LoveLetterEnv(AgentRandom(args.seed), args.seed)
state = env.reset()
shared_model = ActorCritic(state.shape[0], env.action_space).type(dtype)
if args.load_name is not None:
shared_model.load_state_dict(torch.load(args.load_name))
shared_model.share_memory()
# train(1,args,shared_model,dtype)
processes = []
p = mp.Process(target=test,
args=(args.num_processes, args, shared_model, dtype))
p.start()
processes.append(p)
if not args.evaluate:
for rank in range(0, args.num_processes):
def test(rank, args, shared_model, dtype):
test_ctr = 0
torch.manual_seed(args.seed + rank)
# set up logger
timestring = str(date.today()) + '_' + \
time.strftime("%Hh-%Mm-%Ss", time.localtime(time.time()))
run_name = args.save_name + '_' + timestring
configure("logs/run_" + run_name, flush_secs=5)
env = LoveLetterEnv(AgentRandom(args.seed + rank), args.seed + rank)
env.seed(args.seed + rank)
state = env.reset()
model = ActorCritic(state.shape[0], env.action_space).type(dtype)
model.eval()
state = torch.from_numpy(state).type(dtype)
reward_sum = 0
max_reward = -99999999
max_winrate = 0
rewards_recent = deque([], 100)
done = True
start_time = time.time()
episode_length = 0
while True:
episode_length += 1
# Sync with the shared model
if done:
model.load_state_dict(shared_model.state_dict())
cx = Variable(torch.zeros(1, 256).type(dtype), volatile=True)
hx = Variable(torch.zeros(1, 256).type(dtype), volatile=True)
else:
cx = Variable(cx.data.type(dtype), volatile=True)
hx = Variable(hx.data.type(dtype), volatile=True)
value, logit, (hx, cx) = model((Variable(state.unsqueeze(0),
volatile=True), (hx, cx)))
prob = F.softmax(logit)
action = prob.max(1)[1].data.cpu().numpy()
state, reward, done, _ = env.step(action[0, 0])
done = done or episode_length >= args.max_episode_length
reward_sum += reward
if done:
rewards_recent.append(reward_sum)
rewards_recent_avg = sum(rewards_recent) / len(rewards_recent)
print(
"{} | Episode Reward {: >4}, Length {: >2} | Avg Reward {:0.2f}"
.format(
time.strftime("%Hh %Mm %Ss",
time.gmtime(time.time() - start_time)),
reward_sum, episode_length, rewards_recent_avg))
# if not stuck or args.evaluate:
log_value('Reward', reward_sum, test_ctr)
log_value('Reward Average', rewards_recent_avg, test_ctr)
log_value('Episode length', episode_length, test_ctr)
if reward_sum >= max_reward:
# pickle.dump(shared_model.state_dict(), open(args.save_name + '_max' + '.p', 'wb'))
path_output = args.save_name + '_max'
torch.save(shared_model.state_dict(), path_output)
path_now = "{}_{}".format(args.save_name,
datetime.datetime.now().isoformat())
torch.save(shared_model.state_dict(), path_now)
max_reward = reward_sum
win_rate_v_random = Arena.compare_agents_float(
lambda seed: AgentA3C(path_output, dtype, seed),
lambda seed: AgentRandom(seed), 800)
msg = " {} | VsRandom: {: >4}%".format(
datetime.datetime.now().strftime("%c"),
round(win_rate_v_random * 100, 2))
print(msg)
log_value('Win Rate vs Random', win_rate_v_random, test_ctr)
if win_rate_v_random > max_winrate:
print("Found superior model at {}".format(
datetime.datetime.now().isoformat()))
torch.save(
shared_model.state_dict(), "{}_{}_best_{}".format(
args.save_name,
datetime.datetime.now().isoformat(),
win_rate_v_random))
max_winrate = win_rate_v_random
reward_sum = 0
episode_length = 0
state = env.reset()
test_ctr += 1
if test_ctr % 10 == 0 and not args.evaluate:
# pickle.dump(shared_model.state_dict(), open(args.save_name + '.p', 'wb'))
torch.save(shared_model.state_dict(), args.save_name)
if not args.evaluate:
time.sleep(60)
elif test_ctr == evaluation_episodes:
# Ensure the environment is closed so we can complete the
# submission
env.close()
# gym.upload('monitor/' + run_name, api_key=api_key)
state = torch.from_numpy(state).type(dtype)
def train(rank, args, shared_model, dtype):
torch.manual_seed(args.seed + rank)
env = LoveLetterEnv(AgentRandom(args.seed + rank), args.seed + rank)
env.seed(args.seed + rank)
state = env.reset()
model = ActorCritic(state.shape[0], env.action_space).type(dtype)
optimizer = optim.Adam(shared_model.parameters(), lr=args.lr)
model.train()
values = []
log_probs = []
state = torch.from_numpy(state).type(dtype)
done = True
episode_length = 0
while True:
episode_length += 1
# Sync with the shared model
model.load_state_dict(shared_model.state_dict())
if done:
cx = Variable(torch.zeros(1, 256).type(dtype))
hx = Variable(torch.zeros(1, 256).type(dtype))
else:
cx = Variable(cx.data.type(dtype))
hx = Variable(hx.data.type(dtype))
values = []
log_probs = []
rewards = []
entropies = []
for step in range(args.num_steps):
value, logit, (hx, cx) = model(
(Variable(state.unsqueeze(0)), (hx, cx)))
prob = F.softmax(logit)
log_prob = F.log_softmax(logit)
entropy = -(log_prob * prob).sum(1)
entropies.append(entropy)
action = prob.multinomial().data
log_prob = log_prob.gather(1, Variable(action))
state, reward, done, _ = env.step(action.cpu().numpy()[0][0])
done = done or episode_length >= args.max_episode_length
if done:
episode_length = 0
state = env.reset()
state = torch.from_numpy(state).type(dtype)
values.append(value)
log_probs.append(log_prob)
rewards.append(reward)
if done:
break
R = torch.zeros(1, 1).type(dtype)
if not done:
value, _, _ = model((Variable(state.unsqueeze(0)), (hx, cx)))
R = value.data
values.append(Variable(R))
policy_loss = 0
value_loss = 0
R = Variable(R)
gae = torch.zeros(1, 1).type(dtype)
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) - args.beta * entropies[i]
optimizer.zero_grad()
(policy_loss + 0.5 * value_loss).backward()
torch.nn.utils.clip_grad_norm(model.parameters(), 40)
ensure_shared_grads(model, shared_model)
optimizer.step()