def main(model_path, rom_args, learner_args, network_args, num_threads, epochs, logdir, save_interval):
# create env
environment = ALEEnvironment(**rom_args)
# create network then load
num_actions = environment.get_num_actions()
input_shape = [learner_args['phi_length']] + environment.get_state_shape()
network = TargetDQN(input_shape, num_actions, 'dqn', **network_args)
network.load(model_path)
# create threads
del learner_args['epsilon_annealing_start']
learner = QThreadLearner(environment, network, {}, **learner_args, epsilon_annealing_start=0.01, testing=True)
# run 100 episodes
reward_list = []
try:
for _ in range(100):
reward = learner.run_episode(environment)
print('Episode: {}. Steps: {}. Reward: {}'.format(_, environment.curr_step_count, reward))
reward_list.append(reward)
except KeyboardInterrupt:
pass
import matplotlib.pyplot as plt
plt.title('Max: {0}, Mean: {1}, Min: {2}'.format(max(reward_list), np.mean(reward_list), min(reward_list)))
plt.plot(reward_list)
plt.show()
return max(reward_list), np.mean(reward_list), min(reward_list)
def main(rom_args, learner_args, network_args, num_threads, epochs, logdir,
save_interval):
# create envs for each thread
environments = [ALEEnvironment(**rom_args) for _ in range(num_threads)]
# create shared network
num_actions = environments[0].get_num_actions()
input_shape = [learner_args['phi_length']
] + environments[0].get_state_shape()
network = NStepA3CLSTM(input_shape, num_actions, **network_args)
# create thread host
thread_host = AsyncThreadHost(network, log_dir=logdir)
# create threads
threads = [
RecurrentThreadLearner(environments[t], network,
thread_host.shared_dict, **learner_args)
for t in range(num_threads)
]
reward_list = thread_host.run_epochs(epochs,
threads,
save_interval=save_interval)
import matplotlib.pyplot as plt
plt.plot([x[1] for x in reward_list], [x[0] for x in reward_list], '.')
plt.savefig(logdir + 'rewards.png')
plt.show()
return max([x[0] for x in reward_list])