def __init__(self, cfg, tetris):
self.num_actions = cfg.MODEL.SIZE_ACTION
self.gamma = cfg.SOLVER.GAMMA
self.BATCH_SIZE = cfg.SOLVER.BATCH_SIZE
transition = namedtuple('Transicion',
('state', 'action', 'next_state', 'reward'))
self.memory = ReplayMemory(cfg.SOLVER.CAPACITY, transition)
self.model = get_model(cfg)
self.target_net = copy.deepcopy(self.model)
self.target_net.load_state_dict(self.model.state_dict())
self.optimizer = optim.Adam(self.model.parameters(), lr=0.001)
self.tetris = tetris
def __init__(
self,
network: nn.Module,
actions: int,
logger: Optional = None,
learning_rate: float = 0.00025,
replay_start_size: int = 50000,
replay_size: int = 1000000,
batch_size: int = 32,
sync_target_step: int = 10000,
update_frequency: int = 4,
gradient_clipping: bool = False,
reward_clipping: bool = True,
gamma: float = 0.99,
epsilon_start: float = 1.0,
epsilon_end: float = 0.1,
epsilon_end_step: int = 1000000,
epsilon_testing: float = 0.05,
training: bool = True,
device: str = 'gpu',
seed: Optional[int] = None
):
"""
Initializes a DQN agent
Args:
network: a neural network to learn the Q-function
actions: number of actions the agent can take
logger: a logger that has a write method which receives scalars and a timestep
learning_rate: the learning rate for the optimizer
replay_start_size: minimum number of samples in memory before optimization starts, is also the
number of time steps taken before reducing epsilon
replay_size: maximum size of the replay buffer
batch_size: number of samples for each parameter update
sync_target_step: number of policy updates before updating the target network parameters
update_frequency: number of time steps between each learning step
gradient_clipping: if True, the gradients are clipped between -1 and 1
reward_clipping: if True, the rewards are clipped between -1 and 1
gamma: the discount factor for the MDP
epsilon_start: value of epsilon at start of training
epsilon_end: value of epsilon at end of training
epsilon_end_step: number of time steps where the epsilon is linearly decayed
epsilon_testing: value of epsilon during testing
training: if True the agent is training if False is testing
device: device to be used in pytorch, either gpu` or `cpu`
seed: the random seed
"""
if seed is not None:
torch.random.manual_seed(seed)
# selecting the device to use
self._device = torch.device("cuda" if torch.cuda.is_available() and device == 'gpu' else "cpu")
print(f"Using {self._device}...")
# creating the target network, eval doesn't do anything since we are not using dropout
self._policy_network = network.to(self._device)
self._target_network = deepcopy(self._policy_network).to(self._device)
self._target_network.eval()
# saving the logger
if logger is not None:
self._logger = logger
# initializing the optimizer and saving some optimization related parameters
self._learning_rate = learning_rate
# self._optimizer = RMSprop(self._policy_network.parameters(), self._learning_rate)
self._optimizer = torch.optim.Adam(self._policy_network.parameters(), lr=0.0000625, eps=0.00015)
# self._optimizer = torch.optim.Adam(self._policy_network.parameters(), lr=0.0000125, eps=0.00015)
self._batch_size = batch_size
self._sync_target_step = sync_target_step
self._update_frequency = update_frequency
self._gradient_clipping = gradient_clipping
self._loss_fn = torch.nn.L1Loss(reduction="none")
self._reward_clipping = reward_clipping
# setting the action space
self._actions = actions
self._num_steps = 0
# setting the replay buffer
self._replay_start_size = replay_start_size
self._replay_size = replay_size
self._memory = ReplayMemory(size=replay_size, seed=seed)
# setting the MDP parameters
self._gamma = gamma
# setting the exploration parameters
self._epsilon_end = epsilon_end
self._epsilon_diff = epsilon_start - epsilon_end
self._epsilon_end_step = epsilon_end_step
self._epsilon_testing = epsilon_testing
self._epsilon = epsilon_start
# setting the training status
self._training = training
self._timestep = None
self._next_timestep = None
# helper method for reshaping the cartpole observation
def reshape(state):
return np.reshape(state, [1, 4])
if __name__ == '__main__':
tf.compat.v1.disable_eager_execution()
max_score = 0
n_episodes = 5000
max_env_steps = 1000
env = gym.make('CartPole-v0')
agent = DQNAgent(env=env,
net=NN(alpha=0.001, decay=0.0001),
memory=ReplayMemory(size=100000))
if max_env_steps is not None:
env._max_episode_steps = max_env_steps
for e in range(n_episodes):
# reset the env
state = reshape(env.reset())
done = False
score = 0
# play until env done
while not done:
action = agent.act(state)
next_state, reward, done, _ = env.step(action)
# env.render()
next_state = reshape(next_state)
def main():
parser = argparse.ArgumentParser(description='Run DQN on Atari Breakout')
parser.add_argument('--env', default='Breakout-v0', help='Atari env name')
parser.add_argument(
'-o', '--output', default='atari-v0', help='Directory to save data to')
parser.add_argument('--seed', default=0, type=int, help='Random seed')
parser.add_argument('--input_shape', nargs=2, type=int, default=None,
help='Input shape')
parser.add_argument('--num_frame', default=4, type=int,
help='Number of frames in a state')
parser.add_argument('--discount', default=0.99, type=float,
help='Discount factor gamma')
parser.add_argument('--online_train_interval', default=4, type=int,
help='Interval to train the online network')
parser.add_argument('--target_reset_interval', default=10000, type=int,
help='Interval to reset the target network')
parser.add_argument('--action_change_interval', default=1, type=int,
help='Interval to change action')
parser.add_argument('--print_loss_interval', default=100, type=int,
help='Interval to print losses')
parser.add_argument('--replay_buffer_size', default=100000, type=int,
help='Replay buffer size')
parser.add_argument('--num_burn_in', default=25000, type=int,
help='Number of samples filled in memory before update')
parser.add_argument('--batch_size', default=32, type=int,
help='How many samples in each minibatch')
parser.add_argument('--learning_rate', default=1e-4, type=float,
help='Learning rate alpha')
parser.add_argument('--explore_prob', default=0.05, type=float,
help='Exploration probability in epsilon-greedy')
parser.add_argument('--decay_prob_start', default=1.0, type=float,
help='Starting probability in linear-decay epsilon-greedy')
parser.add_argument('--decay_prob_end', default=0.1, type=float,
help='Ending probability in linear-decay epsilon-greedy')
parser.add_argument('--decay_steps', default=1000000, type=int,
help='Decay steps in linear-decay epsilon-greedy')
parser.add_argument('--num_train', default=5000000, type=int,
help='Number of training sampled interactions with the environment')
parser.add_argument('--max_episode_length', default=999999, type=int,
help='Maximum length of an episode')
parser.add_argument('--save_interval', default=100000, type=int,
help='Interval to save weights and memory')
parser.add_argument('--model_name', default='dqn', type=str,
help='Model name')
parser.add_argument('--eval_interval', default=10000, type=int,
help='Evaluation interval')
parser.add_argument('--eval_episodes', default=20, type=int,
help='Number of episodes in evaluation')
parser.add_argument('--double_q', default=False, type=bool,
help='Invoke double Q net')
parser.add_argument('--do_render', default=False, type=bool,
help='Do rendering or not')
parser.add_argument('--read_weights', default=None, type=str,
help='Read weights from file')
parser.add_argument('--read_memory', default=None, type=str,
help='Read memory from file')
args = parser.parse_args()
print '########## All arguments ##########:', args
args.input_shape = tuple(args.input_shape)
args.output = get_output_folder(args.output, args.env)
env = gym.make(args.env)
num_actions = env.action_space.n
opt_adam = Adam(lr=args.learning_rate)
model_online = create_model(args.num_frame, args.input_shape,
num_actions, model_name=args.model_name)
model_target = create_model(args.num_frame, args.input_shape,
num_actions, model_name=args.model_name)
q_network = {'online': model_online, 'target': model_target}
preproc = AtariPreprocessor(args.input_shape)
memory = ReplayMemory(args.replay_buffer_size, args.num_frame)
policy_random = UniformRandomPolicy(num_actions)
policy_train = LinearDecayGreedyEpsilonPolicy(args.decay_prob_start,
args.decay_prob_end,
args.decay_steps)
policy_eval = GreedyEpsilonPolicy(args.explore_prob)
policy = {'random': policy_random, 'train': policy_train, 'eval': policy_eval}
agent = DQNAgent(num_actions, q_network, preproc, memory, policy, args)
agent.compile([mean_huber_loss, null_loss], opt_adam)
if args.read_weights is not None:
agent.q_network['online'].load_weights(args.read_weights)
if args.read_memory is not None:
with open(args.read_memory, 'rb') as save_memory:
agent.memory = pickle.load(save_memory)
print '########## training #############'
agent.fit(env)
import gym
from dqn.bots import AtariBot
from dqn.policy import DDQNPolicy
from dqn.memory import ReplayMemory
GAME = 'Breakout-v0'
#TODO List params to tune here, eventually migrate this to a readme
if __name__ == "__main__":
policy = DDQNPolicy()
memory = ReplayMemory()
game = gym.make(GAME)
game.ale.setInt(b'frame_skip', 4)
robot = AtariBot(policy=policy, memory=memory)
robot.train(game=game, ckpt_dir="models")