def __init__(self, env, params, net=None, reward=[], loss=[]):
self.params = params
self.r_sums = reward
self.l_means = loss
# PARAMS
self.gamma = params["gamma"]
self.freq_copy = params["freq_copy"]
self.tau = params["max_tau"]
self.tau_decay = params["tau_decay"]
self.min_tau = params["min_tau"]
self.exploration = params["exploration"]
self.sigma = params["sigma"]
self.alpha = params["alpha"]
self.m = params["m"]
self.frame_skip = params["frame_skip"]
self.target_update_strategy = params["target_update_strategy"]
self.batch_size = params["batch_size"]
self.cuda = False
# NEURAL NETWORK
self.n_action = env.action_space.n
self.net = QModel(self.n_action)
if net is not None:
self.net.load_state_dict(net)
self.target = copy.deepcopy(self.net)
self.optimizer = params["optimizer"](self.net.parameters(),
lr=self.sigma)
self.criterion = params["criterion"]()
self.buff = Buffer(params["buffer_size"])
self.env = wrappers.AtariPreprocessing(env,
frame_skip=self.frame_skip,
screen_size=84,
grayscale_obs=True,
scale_obs=True)
self.env = wrappers.FrameStack(self.env, self.m)
def __init__(self,
variant="dueling",
use_gpus=False,
learning_rate=0.00025,
headless=False,
epoch_length=50000):
super().__init__()
#self.env = gym.make("Pong-v0")
self.headless = headless
self.env = gym.make("Pong-v0")
self.env = wrappers.FrameStack(
wrappers.ResizeObservation(wrappers.GrayScaleObservation(self.env),
84), 4)
if self.headless:
self.virtual_display = Display(visible=0, size=(1400, 900))
self.virtual_display.start()
self.env = wrappers.Monitor(self.env,
"./video/Pong-v0",
video_callable=lambda episode_id: True,
force=True)
self.observation = self.env.reset()
self.epsilon = 1
self.discount_factor = 0.99
self.network = DuelingGameNet(
self.env.action_space.n,
use_gpus) if variant == "dueling" else StandardGameNet(
self.env.action_space.n, use_gpus)
self.target_network = DuelingGameNet(
self.env.action_space.n,
use_gpus) if variant == "dueling" else StandardGameNet(
self.env.action_space.n, use_gpus)
self.target_update_rate = 1000
self.reward = 0
self.game_reward = 0
self.games = 0
self.epoch_length = epoch_length
self.learning_rate = learning_rate
self.steps_to_train = 4
self.last_ten = Memory(10)
self.variant = variant
print("Using environment:", env_name)
if 'StarGunner' in env_name:
use_conv = True
env = gym.wrappers.AtariPreprocessing(env,
frame_skip=4,
grayscale_obs=False,
screen_size=84)
else:
use_conv = False
if 'StarGunner' in env_name:
#env = wrappers.EpisodicLifeEnv(env)
env = wrappers.GrayScaleObservation(env)
env = ResizeObservationKeepDims(env, (84, 84))
env = wrappers.FrameStack(env, 4, lz4_compress=True)
if calc_rolling_avg:
env = RollingMeanReturn(env, window=100)
env = RecordInfo(env, rolling_avgs_file, ["episode"], overwritefile=False)
import DQN
import GymTrainer as gt
# try to resume training, or start new training example
try:
model = DQN.DQNmodel(env.action_space.n,
folder, {"minutes": 1},
use_convolutions=use_conv)
replay_mem_file = model.load(backup_folder=folder)
assert replay_mem_file
with open(os.path.join(eval_args.path, 'args.json'), 'r') as f:
train_args = ArgsStruct(**json.load(f))
env_name = eval_args.env if eval_args.env is not None else train_args.env
episodes = eval_args.episodes if eval_args.episodes is not None else train_args.optimize_freq
timestamp = time.strftime("%Y-%m-%d-%H%M")
log_path = os.path.join(exp_dir, f'eval-{timestamp}.log')
logging.basicConfig(filename=log_path, level=logging.INFO)
logging.getLogger('').addHandler(logging.StreamHandler())
if env_name == 'pong':
env_id = 'PongNoFrameskip-v0'
env = gym.make(env_id)
else:
raise NotImplementedError(env_name)
device = torch.device('cuda:0' if not eval_args.no_cuda
and torch.cuda.is_available() else 'cpu')
env = wrappers.FrameStack(wrappers.AtariPreprocessing(env), num_stack=4)
model = load_agent(train_args, env).to(device)
model.load_state_dict(
torch.load(os.path.join(exp_dir, 'checkpoint_best.pt')))
eval_res = evaluate(model, env, train_args, device, episodes)
logging.info(pformat(eval_res))
args = parser.parse_args()
env = gym.make(args.env_id)
env.spec.id += " NoFrameskip"
# You provide the directory to write to (can be an existing
# directory, including one with existing data -- all monitor files
# will be namespaced). You can also dump to a tempdir if you'd
# like: tempfile.mkdtemp().
outdir = '/tmp/random-agent-results'
env = wrappers.Monitor(env, directory=outdir, force=True)
env = wrappers.AtariPreprocessing(env,
screen_size=84,
frame_skip=4,
grayscale_obs=True)
env = wrappers.FrameStack(env, 4)
env.seed(0)
neural_network = ConvNet()
target_neural_network = copy.deepcopy(neural_network)
print(list(neural_network.parameters()))
criterion = nn.MSELoss()
optim = torch.optim.SGD(neural_network.parameters(),
lr=LEARNING_RATE,
momentum=MOMENTUM)
optim.zero_grad()
reward = 0
buffer = deque(maxlen=10000)
done = False
reward_evolution = []
epsilon_greedy_evolution = []
# You can set the level to logger.DEBUG or logger.WARN if you
# want to change the amount of output.
logger.set_level(logger.WARN)
env = gym.make(args.env_id)
outdir = './videos/' + module
env = wrappers.Monitor(env, directory=outdir, force=True)
# You provide the directory to write to (can be an existing
# directory, including one with existing data -- all monitor files
# will be namespaced). You can also dump to a tempdir if you'd
# like: tempfile.mkdtemp().
env.seed(0)
preproc = wrappers.AtariPreprocessing(env)
preproc = wrappers.FrameStack(preproc, 4)
neur = CNN(4, env.action_space.n)
neur.load_state_dict(torch.load('./trained_networks/' + module + '.n'))
neur.eval()
agent = AgentAtari(4, env.action_space.n, False, neur)
episode_count = 5
reward = 0
done = False
reward_accumulee = 0
for _ in range(episode_count):
ob = preproc.reset()
def train(args: argparse.Namespace, env: gym.Env, exp_dir: str):
seed = args.seed
torch.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(seed)
device = torch.device(
'cuda:0' if not args.no_cuda and torch.cuda.is_available() else 'cpu')
env = wrappers.FrameStack(wrappers.AtariPreprocessing(env),
num_stack=args.stacked_frames)
writer = SummaryWriter(log_dir=exp_dir)
with open(os.path.join(exp_dir, 'args.json'), 'w') as f:
json.dump(args.__dict__, f, indent=2)
logging.info(args)
n_actions = env.action_space.n
current_model = load_agent(args, env).to(device)
current_model.eval()
target_model = load_agent(args, env).to(device)
target_model.eval()
if args.curiosity:
curiosity = load_icm(args, env).to(device)
curiosity.eval()
target_model.load_state_dict(
current_model.state_dict()) # Sync/update target model
# rms-prop? https://www.reddit.com/r/reinforcementlearning/comments/ei9p3y/using_rmsprop_over_adam/
if args.optimizer == 'adam':
optimizer = optim.Adam(current_model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
if args.curiosity:
curiosity_optimizer = optim.Adam(curiosity.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
else:
logging.error('Optimizer not implemented')
raise NotImplementedError()
logging.info(current_model)
if args.curiosity:
logging.info(curiosity)
n_params = sum(p.numel() for p in current_model.parameters()
if p.requires_grad)
logging.info(f'Training {n_params} parameters')
if args.curiosity:
n_params = sum(p.numel() for p in curiosity.parameters()
if p.requires_grad)
logging.info(f'Training {n_params} parameters')
criterion = nn.SmoothL1Loss if args.criterion == 'huber' else None
if criterion is None:
raise NotImplementedError(args.criterion)
buffer = ReplayBuffer(capacity=args.replay_size, seed=args.seed)
best_mean_reward = env.reward_range[0]
updates_without_improvement = 0
# Adapted from Mario Martin's Notebook
epsilon_start = 1.0
epsilon_final = 0.01
epsilon_decay = 10000
epsilon_by_episode = lambda e: epsilon_final + (
epsilon_start - epsilon_final) * math.exp(-1. * e / epsilon_decay)
if args.curiosity:
epsilon_by_episode = lambda e: 0.0 # No epsilon needed if curiosity is used
t0 = time.time()
all_rewards = []
all_steps = []
all_mean_rewards = []
all_mean_steps = []
episode_set_rewards = 0.0
episode_set_curiosity_rewards = 0.0
episode_set_steps = 0
updates = 0
optimizations = 0
initial_counter = 0
for episode in range(args.episodes):
state = env.reset()
episode_reward = 0.0
episode_curiosity_reward = 0.0
steps = 0
epsilon = epsilon_by_episode(episode)
while True:
current_model.eval()
if args.curiosity:
curiosity.eval()
action = current_model.act(
torch.tensor(transform(
state.__array__())).unsqueeze(0).to(device), epsilon,
torch.rand(1)[0].to(device),
torch.randint(0, n_actions, (1, ))[0].to(device))
current_model.train()
next_state, reward, done, _ = env.step(action)
episode_reward += reward
curiosity_reward = None
if args.curiosity:
with torch.no_grad():
curiosity_reward, _ = \
curiosity(torch.tensor(transform(state.__array__())).unsqueeze(0).to(device),
torch.tensor(transform(next_state.__array__())).unsqueeze(0).to(device),
torch.tensor([action]).long().to(device))
episode_curiosity_reward += curiosity_reward
buffer.push(
LazyTransition(
state, action, next_state, reward, done,
curiosity_reward.cpu().numpy()
if curiosity_reward is not None else None))
if done:
initial_counter += 1
writer.add_scalar('Reward/train', episode_reward, episode + 1)
writer.add_scalar('Steps/train', steps, episode + 1)
writer.add_scalar('Epsilon/train', epsilon, episode + 1)
all_rewards.append(episode_reward)
all_steps.append(steps)
episode_set_rewards += episode_reward
episode_set_steps += steps
if args.curiosity:
writer.add_scalar('Curiosity/train',
episode_curiosity_reward, episode + 1)
episode_set_curiosity_rewards += episode_curiosity_reward
state = next_state
steps += 1
if args.render:
env.render()
if done:
logging.info(
f'Finished episode {episode+1} with reward = {episode_reward:.2f} | steps = {steps+1} | '
f'epsilon = {epsilon:.2f}')
if args.curiosity:
logging.info(f'curiosity = {curiosity_reward:.2f}')
break
if buffer.full and (
episode + 1
) % args.optimize_freq == 0: # len(buffer) >= args.batch_size:
transitions = buffer.sample(args.batch_size)
if not args.curiosity:
q_loss, _ = optimize(transitions, current_model, target_model,
optimizer, device, epsilon,
args.criterion)
else:
q_loss, curiosity_loss = optimize(transitions, current_model,
target_model, optimizer,
device, epsilon,
args.criterion, curiosity,
curiosity_optimizer)
denominator = args.optimize_freq - 1 if optimizations > 0 else initial_counter
mean_episode_set_rewards = episode_set_rewards / denominator
mean_episode_set_steps = episode_set_steps / denominator
writer.add_scalar('Mean-Reward/train', mean_episode_set_rewards,
optimizations + 1)
writer.add_scalar('Mean-Steps/train', mean_episode_set_steps,
optimizations + 1)
writer.add_scalar('Q-Loss/train', q_loss, optimizations + 1)
if args.curiosity:
writer.add_scalar('Curiosity-Loss/train', curiosity_loss,
optimizations + 1)
all_mean_rewards.append(mean_episode_set_rewards)
all_mean_steps.append(mean_episode_set_steps)
episode_set_rewards = 0.0
episode_set_steps = 0
torch.save(current_model.state_dict(),
os.path.join(exp_dir, 'checkpoint_last.pt'))
if args.curiosity:
torch.save(
curiosity.state_dict(),
os.path.join(exp_dir, 'curiosity_checkpoint_last.pt'))
logging.info(f'Optimized model ({optimizations+1} optimizations)')
optimizations += 1
if mean_episode_set_rewards > best_mean_reward:
updates_without_improvement = 0
best_mean_reward = mean_episode_set_rewards
torch.save(current_model.state_dict(),
os.path.join(exp_dir, 'checkpoint_best.pt'))
logging.info(f'NEW: Best mean reward: {best_mean_reward:.2f}')
if best_mean_reward == env.reward_range[1]:
logging.info('Reached max reward')
break
else:
updates_without_improvement += 1
logging.info(f'Best mean reward: {best_mean_reward:.2f}')
if args.early_stop != -1 and updates_without_improvement == args.early_stop:
break
logging.info(
f'{updates_without_improvement} updates without improvement')
if buffer.full and (episode + 1) % args.update_target_freq == 0:
target_model.load_state_dict(current_model.state_dict())
logging.info(f'Updated target model (updates {updates+1})')
updates += 1
t1 = time.time()
logging.info(f'Finished training in {t1-t0:.1f}s')
if args.render:
env.close()
model = load_agent(args, env).to(device)
model.load_state_dict(torch.load('checkpoint_best.pt'))
eval_res = evaluate(model, env, args, device, episodes=args.optimize_freq)
logging.info(pformat(eval_res))