def start_threads(self):
# max number of episodes
max_eps = 1e6
envs = []
# create 1 local enviroment for each thread
for _ in range(NUM_THREADS):
_env = gym_super_mario_bros.make(env_name)
_env = JoypadSpace(_env, SIMPLE_MOVEMENT)
env = atari_wrapper.wrap_dqn(_env)
envs.append(env)
# create the threads and assign them their enviroment and exploration rate
threads = []
for i in range(NUM_THREADS):
thread = threading.Thread(
target=train_thread,
daemon=True,
args=(self, max_eps, envs[i],
agent.discount_rate, self.optimizer, stats,
AnnealingVariable(.7, 1e-20, 10000), i))
threads.append(thread)
# starts the threads
for t in threads:
print("STARTING")
t.start()
time.sleep(0.5)
try:
[t.join() for t in threads] # wait for threads to finish
except KeyboardInterrupt:
print("Exiting threads!")
action = self.epsilon_greedy_policy(state, exploration_rate)
next_state, reward, done, _ = env.step(action)
episode_reward += reward
state = next_state
rewards_arr[episode] = episode_reward
stats(self, episode_reward)
print(episode_reward)
stats.save_stats()
return rewards_arr
def save_weights(self):
print("PRINT")
self.model.save_weights("PongDQNWeights.h5")
def save_model(self):
self.model.save("DqnPongModel.h5")
def restore_weights(self):
print("Restoring model weights Pong")
self.model.load_weights("DqnPongModel.h5")
stats = Stats()
_env = gym.make('PongNoFrameskip-v4')
env = atari_wrapper.wrap_dqn(_env)
agent = DQNAgent()
agent.init_priority_replay(50000)
agent.train_model(max_steps=50e6, stats=stats)
env.close()
env.render()
# time.sleep(0.05)
action = self.pick_action(state, exploration_rate)
next_state, reward, done, _ = env.step(action)
episode_reward += reward
state = next_state
if callable(stats):
stats(self, episode_reward)
rewards_arr[episode] = episode_reward
print(episode_reward)
stats.save_stats()
return rewards_arr
def restore_weights(self):
self.global_network.load_weights('A3CPong.h5')
test_env = gym.make(env_name)
test_env = atari_wrapper.wrap_dqn(test_env)
NUM_ACTIONS = test_env.action_space.n
OBS_SPACE = test_env.observation_space.shape[0]
state = test_env.reset()
state = np.expand_dims(state, axis=0)
stats = Stats()
agent = A3CAgent()
agent.play(test_env, stats)
test_env.close()