def setup(config: str, load_file: str) -> DQNAgent:
if config:
loader = AgentLoader(config, num_actions=num_actions, num_inputs=num_inputs)
agent = loader.load()
else:
agent = DQNAgent(num_actions=num_actions, num_inputs=num_inputs)
if load_file:
print(f'Loading "{load_file}"...')
agent.load(load_file)
return agent
output = self.l5(h4)
return output
net = Linear()
print('Initializing the learner...')
learner = Learner(settings)
learner.load_net(net)
print('Initializing the agent framework...')
agent = DQNAgent(settings)
print('Training...')
agent.train(learner, memory, simulator)
print('Loading the net...')
learner = agent.load(settings['save_dir'] + '/learner_final.p')
ind_max = learner.val_rewards.index(max(learner.val_rewards))
ind_net = settings['initial_exploration'] + ind_max * settings['eval_every']
agent.load_net(learner, settings['save_dir'] + '/net_%d.p' % int(ind_net))
np.random.seed(settings["seed_general"])
print('Evaluating DQN agent...')
print('(reward, MSE loss, mean Q-value, episodes - NA, time)')
reward, MSE_loss, mean_Q_value, episodes, time, paths, actions, rewards = agent.evaluate(
learner, simulator, 50000)
print(reward, MSE_loss, mean_Q_value, episodes, time)