def __init__(self, size):
self.size = size
self.memory = deque(maxlen=self.size)
def update(self, SARS):
self.memory.append(SARS)
def sample(self, batch_size):
return zip(*random.sample(self.memory, batch_size))
r_memory = ReplayMemory(memory_size)
agent = DQN(12, 12, 16)
target = DQN(12, 12, 16)
target.load_state_dict(agent.state_dict())
optimizer = Adam(agent.parameters())
def update_target():
if len(r_memory.memory) < batch_size:
return
observation, action, reward, observation_next, done = r_memory.sample(
batch_size)
observations = torch.cat(observation)
observation_next = torch.cat(observation_next)
actions = index_action(torch.LongTensor(action))
rewards = torch.LongTensor(reward)
done = torch.FloatTensor(done)
q_values = agent(observations)
p_q_values_next = agent(observation_next)
q_values_next = target(observation_next)
loss.backward()
optimizer.step()
return loss.data[0]
def save_checkpoint(state, filename):
torch.save(state, filename)
#def load_model():
#torch.load(
if __name__ == '__main__':
model = DQN()
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
criterion = nn.CrossEntropyLoss()
# Load data
data = np.load('training_data.npy')
X_train = np.stack(data[:, 0], axis=0).reshape(
(len(data), 1, len(data[0][0]), len(data[0][0][0])))
Y_train = data[:, 3]
# Training loop
for epoch in range(100):
# Randomize and batch training data
batchsize = 8
# Randomly shuffle each epoch
np.random.shuffle(X_train)
np.random.shuffle(Y_train)