def __init__(self, env_id="PongNoFrameskip-v4"):
super(CnnDQN, self).__init__()
self.env_id = env_id
self.env = make_atari(self.env_id)
self.env = wrap_deepmind(self.env)
self.env = wrap_pytorch(self.env)
self.input_shape = self.env.observation_space.shape
self.num_actions = self.env.action_space.n
self.features = nn.Sequential(
nn.Conv2d(self.env.observation_space.shape[0], 32, kernel_size=8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, kernel_size=4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, kernel_size=3, stride=1),
nn.ReLU()
)
self.fc = nn.Sequential(
nn.Linear(self.feature_size(), 512),
nn.ReLU(),
nn.Linear(512, self.num_actions)
)
self.optimizer = optim.Adam(self.parameters(), lr=0.00001)
self.replay_buffer = ReplayBuffer2(100000)
self.losses = []
def __init__(self,
env_id="PongNoFrameskip-v4",
num_atoms=51,
Vmin=-10,
Vmax=10):
self.num_atoms = num_atoms
self.Vmin = Vmin
self.Vmax = Vmax
self.env_id = env_id
self.env = make_atari(self.env_id)
self.env = wrap_deepmind(self.env)
self.env = wrap_pytorch(self.env)
self.current_model = TinyRainbowCnnDQN(
self.env.observation_space.shape, self.env.action_space.n,
num_atoms, Vmin, Vmax)
self.target_model = TinyRainbowCnnDQN(self.env.observation_space.shape,
self.env.action_space.n,
num_atoms, Vmin, Vmax)
if USE_CUDA:
self.current_model = self.current_model.cuda()
self.target_model = self.target_model.cuda()
self.optimizer = optim.Adam(self.current_model.parameters(), lr=0.0001)
self.update_target(self.current_model, self.target_model)
self.replay_buffer = ReplayBuffer(100000)
self.losses = []
def __init__(self, env_id="PongNoFrameskip-v4"):
self.env_id = env_id
self.env = make_atari(self.env_id)
self.env = wrap_deepmind(self.env)
self.env = wrap_pytorch(self.env)
self.current_model = CnnDQN(self.env.observation_space.shape, self.env.action_space.n)
self.target_model = CnnDQN(self.env.observation_space.shape, self.env.action_space.n)
if USE_CUDA:
self.current_model = self.current_model.cuda()
self.target_model = self.target_model.cuda()
self.optimizer = optim.Adam(self.current_model.parameters(), lr=0.0001)
self.replay_buffer = NaivePrioritizedBuffer(100000)
self.update_target(self.current_model, self.target_model)
self.losses = []
def __init__(self, env_id= "PongNoFrameskip-v4", replay_buffer_size=100000):
self.env_id = env_id
self.env = make_atari(self.env_id)
self.env = wrap_deepmind(self.env)
self.env = wrap_pytorch(self.env)
self.current_model = CnnDQN()
self.target_model = CnnDQN()
if torch.cuda.is_available():
self.current_model = self.current_model.cuda()
self.target_model = self.target_model.cuda()
self.optimizer = optim.Adam(self.current_model.parameters(), lr=0.0001)
self.replay_buffer = ReplayBuffer2(replay_buffer_size)
self.losses = []
self.target_model.load_state_dict(self.current_model.state_dict())
def __init__(self, env_id="PongNoFrameskip-v4"):
self.env_id = env_id
self.env = make_atari(self.env_id)
self.env = wrap_deepmind(self.env)
self.env = wrap_pytorch(self.env)
self.current_model = TinyDuelingCnnDQN(
self.env.observation_space.shape, self.env.action_space.n)
self.target_model = TinyDuelingCnnDQN(self.env.observation_space.shape,
self.env.action_space.n)
if torch.cuda.is_available():
self.current_model = self.current_model.cuda()
self.target_model = self.target_model.cuda()
self.optimizer = optim.Adam(self.current_model.parameters(), lr=0.0001)
self.replay_buffer = ReplayBuffer2(100000)
self.update_target(self.current_model, self.target_model)
self.losses = []