load_pretrained_network = False
print('Initialising VizDoom...')
config_path = 'scenarios/basic.cfg'
actor_path = 'models/basic_actor.hd5'
critic_path = 'models/basic_critic.hd5'
reward_table = OrderedDict({'FRAGCOUNT': 1})
resolution = (90, 60)
doom = VizdoomWrapper(config_path=config_path,
reward_table=reward_table,
frame_resolution=resolution,
show_mode=False,
frame_stack=1)
print('Initialising Doomguy...')
doomguy = PPODoom(doom.get_state_size(),
doom.get_action_size(),
actor_path=actor_path,
critic_path=critic_path)
if load_pretrained_network:
doomguy.load_model()
for epoch in range(epochs):
print('\nEpoch {}\n-------'.format(epoch + 1))
print('\nTraining...')
doom.new_game()
train_scores = []
train_losses = []
prev_variables = []
tqdm_bar = tqdm.trange(training_episodes_per_epoch, leave=False)
model_path = 'models/health.hd5'
reward_table = OrderedDict({})
resolution = (84, 84)
episodes = 500
gamma = 0.99
learning_rate = 0.0002
print('Initialising Doom...')
doom = VizdoomWrapper(config_path=config_path,
reward_table=reward_table,
frame_resolution=resolution,
show_mode=False,
frame_stack=4)
doomguy = PolicyGradientAgent(doom.get_state_size(),
doom.get_action_size(),
learning_rate,
gamma,
save_path=model_path)
if load_pretrained_network:
doomguy.load_model()
if train_network:
for episode in range(episodes):
print('Episode', episode)
doom.new_game()
done = False
step = 0
episodes_to_watch = 2
print('Initialising VizDoom...')
config_path = 'scenarios/basic.cfg'
actor_path = 'models/defend_the_center_actor.hd5'
critic_path = 'models/defend_the_center_critic.hd5'
reward_table = OrderedDict({'FRAGCOUNT': 1})
resolution = (90, 60)
doom = VizdoomWrapper(config_path=config_path,
reward_table=reward_table,
frame_resolution=resolution,
show_mode=True,
frame_stack=1)
print('Initialising Doomguy...')
doomguy = BaseQDoom(doom.get_state_size(), doom.get_action_size())
doomguy.load_model()
for _ in range(episodes_to_watch):
done = False
doom.new_game()
while not done:
state = doom.get_current_state()
best_action = doomguy.act(state)
done = doom.set_action(best_action)
# Sleep between episodes
sleep(1.0)
score = doom.get_total_reward()
print('Total score: {}'.format(score))