def test_generator(self):
env = TradeEnv()
board = env.new_sec()
print(board.current_time)
print(board.q_value)
board = env.new_sec()
print(board.q_value)
print(board.current_time)
board = env.new_sec()
print(board.q_value)
print(board.current_time)
def __init__(self, buffer_size=BUFFER_SIZE):
self.buffer_size = buffer_size
self.experiences = deque(maxlen=self.buffer_size)
self.q_values = deque(maxlen=self.buffer_size)
self.reward = 0
self.start_time = 0
self.end_time = 0
self.loss = None
self.total_reward = 0
self.duration = 0
self.episode_no = 0
self.env = TradeEnv()
def test_action1(self):
env = TradeEnv()
print(env.board.current_time)
print(env.q_value)
env.step(ACTION.NOP)
print(env.board.current_time)
print(env.q_value)
env.step(ACTION.SELL_NOW)
print(env.board.current_time)
print(env.q_value)
env.step(ACTION.BUY_NOW)
def test_buy(self):
env = TradeEnv()
env.action_buy()
def test_create(self):
env = TradeEnv()
self.assertTrue(env is not None)
def test_action_sell(self):
env = TradeEnv()
env.step(ACTION.SELL)
print('1')
env.step(ACTION.BUY)
print('2')
env.step(ACTION.SELL)
print('3')
env.step(ACTION.BUY)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
env.step(ACTION.SELL)
def test_action2(self):
env = TradeEnv()
env.step(ACTION.SELL)
env.step(ACTION.BUY)
def test_close(self):
env = TradeEnv()
env.close()
def test_seed(self):
env = TradeEnv()
env.seed()
def test_step(self):
env = TradeEnv()
env.step(1)
def test_render(self):
env = TradeEnv()
env.render()
def test_reset(self):
env = TradeEnv()
env.reset()
def test_sell_now(self):
env = TradeEnv()
env.action_sell_now()
def test_buy_now(self):
env = TradeEnv()
env.action_buy_now()
def test_sell(self):
env = TradeEnv()
env.action_sell()
def test_new_episode(self):
env = TradeEnv()
env.new_episode()
def _init():
e = TradeEnv()
e.seed(seed + rank)
return e
class Trainer:
def __init__(self, buffer_size=BUFFER_SIZE):
self.buffer_size = buffer_size
self.experiences = deque(maxlen=self.buffer_size)
self.q_values = deque(maxlen=self.buffer_size)
self.reward = 0
self.start_time = 0
self.end_time = 0
self.loss = None
self.total_reward = 0
self.duration = 0
self.episode_no = 0
self.env = TradeEnv()
def train(self, episode=NUM_OF_EPISODE):
for i in range(episode):
last_q_time = 0
self.episode_begin(i, None)
self.env.reset()
n_state, reward, done, info = self.env.step(ACTION.NOP)
s = n_state
while True:
if not self.env.q_value:
break
if last_q_time != self.env.q_value.time:
last_q_time = self.env.q_value.time
a = self.env.q_value.get_best_action()
if a != ACTION.NOP:
print(self.env.q_value)
n_state, reward, done, info = self.env.step(a)
self.reward += reward
if (n_state is not None) and (self.env.q_value is not None):
q = QState(n_state, self.env.q_value)
self.q_values.append(q)
s = n_state
if done:
break
self.episode_end(i, s)
np.savez_compressed('/tmp/q_values.npz', self.q_values)
states = np.array([q.s for q in self.q_values])
q_values = np.array([q.q.to_array() for q in self.q_values])
states = states.reshape(states.shape)
q_values = q_values.reshape(q_values.shape)
np.savez_compressed('/tmp/q_stats.npz', s=states, q=q_values)
def learning(self):
npz = np.load('/tmp/q_stats.npz')
states = npz['s']
q_values = npz['q']
print('stateshape', states.shape)
print('qvalueshape', q_values.shape)
#reg = ImageRegressor(output_dim=5, seed=12314, max_trials=3)
reg = self.create_image_regressor()
reg.fit(states, q_values, validation_split=0.2, epochs=2)
model = reg.export_model()
print(type(model))
model.save('./auto_model.hd5')
def create_image_regressor(self):
input_node = ak.ImageInput()
output_node = ak.ConvBlock()(input_node)
output_node = ak.DenseBlock()(output_node)
output_node = ak.RegressionHead()(output_node)
reg = ak.AutoModel(inputs=input_node,
outputs=output_node,
max_trials=10)
return reg
def episode_begin(self, i: int, s):
pass
def episode_end(self, i: int, s):
self.duration = float(self.end_time) - float(self.start_time)
self.total_reward += self.reward
s = '<- EPISODE END ({:5d}) TOTAL{:5.2f}'.format(i, self.total_reward)
self.episode_no += 1