def test_init_memory(self):
model = Sequential()
when(model).compile(loss='mean_squared_error',
optimizer='Adam').thenReturn()
environment = mock(
{
'observation_space': Discrete(8),
'action_space': Discrete(3)
}, FL)
when(environment).reset().thenReturn(0)
#when(environment).step(...).thenReturn((1, 10, True))
expect(environment, times=2).step(...).thenReturn((1, 10, True))
configuration = mock(
{
'model': model,
'memory_size': 2,
'nb_iterations': 0,
'training_params': [],
'plot_training': False
}, Configuration)
test_Agent = DQN(environment, configuration)
verify(environment, times=2).step(...)
def __init__(
self,
state_size,
n_actions,
args,
device=torch.device("cuda" if torch.cuda.is_available() else "cpu")):
super().__init__(state_size, n_actions, args, device=device)
self.target_net = DQN(state_size, n_actions,
layers=self.layers).to(self.device)
self.target_net.load_state_dict(self.policy_net.state_dict())
self.target_net.eval()
def __init__(
self,
state_size,
n_actions,
args,
device=torch.device("cuda" if torch.cuda.is_available() else "cpu")):
self.device = device
# Exploration / Exploitation params.
self.steps_done = 0
self.eps_threshold = 1
self.eps_start = args.eps_start
self.eps_end = args.eps_end
self.eps_decay = args.eps_decay
# RL params
self.target_update = args.target_update
self.discount = args.discount
# Env params
self.n_actions = n_actions
self.state_size = state_size
# Deep q networks params
self.layers = args.layers
self.batch_size = args.batch_size
self.policy_net = DQN(state_size, n_actions,
layers=self.layers).to(self.device).float()
self.target_net = None
self.grad_clip = args.grad_clip
if str(args.optimizer).lower() == 'adam':
self.optimizer = optim.Adam(self.policy_net.parameters())
if str(args.optimizer).lower() == 'rmsprop':
self.optimizer = optim.RMSprop(self.policy_net.parameters())
else:
raise NotImplementedError
self.memory = ReplayMemory(args.replay_size)
# Performance buffers.
self.rewards_list = []