def __init__(self, config):
config.model_type = Config.MODEL_TYPE_CONV2D
print(config.device)
self.config = config
self.env = wrap_deepmind(make_atari(config.env), frame_stack=True)
self.num_states = self.env.observation_space.shape[-1]
self.num_actions = self.env.action_space.n
self.agent = Agent(config, self.num_states, self.num_actions, self.config.num_atoms)
self.total_step = np.zeros(100)
self.data_path = config.data_path
if self.data_path != Config.DATA_PATH_DEFAULT:
self.agent.load_model()
def __init__(self, config):
config.model_type = Config.MODEL_TYPE_CONV2D
print(config.device)
self.config = config
self.env = wrap_deepmind(make_atari(config.env), frame_stack=True)
self.num_states = self.env.observation_space.shape[-1]
self.num_actions = self.env.action_space.n
self.agent = Agent(config, self.num_states, self.num_actions,
self.config.num_atoms)
self.total_step = np.zeros(100)
self.data_path = config.data_path
if self.data_path != Config.DATA_PATH_DEFAULT:
self.agent.load_model()
def get_env():
benchmark = gym.benchmark_spec('Atari40M')
# Change the index to select a different game.
task = benchmark.tasks[3]
env_id = task.env_id
env = gym.make(env_id)
env.seed(0)
expt_dir = '/tmp/hw3_vid_dir2/'
# env = wrappers.Monitor(env, osp.join(expt_dir, "gym"), force=True)
env = utils.wrap_deepmind(env)
return env
log_freq=args.log_freq,
mode=args.mode)
plotter = Plotter(save_dirs=save_dirs,
plot_types=[
'avg_scores_ep', 'avg_scores_ts', 'avg_scores_100_ep',
'avg_scores_100_ts', 'scores_ep', 'scores_ts',
'high_scores_ep', 'high_scores_ts', 'low_scores_ep',
'low_scores_ts', 'avg_loss_ep', 'avg_acc_ep',
'timesteps_ep'
],
interval_types=['overall', 'window'],
plot_freq=args.plot_freq,
mode=args.mode)
env = make_atari(ENV_GYM)
env = wrap_deepmind(env, frame_stack=True, scale=False)
if args.mode == 'train':
agent = DDQNLearner(
env=env,
save_dirs=save_dirs,
save_freq=args.save_freq,
gamma=args.gamma,
batch_size=args.batch_size,
learning_rate=args.learning_rate,
buffer_size=args.buffer_size,
learn_start=args.learn_start,
target_network_update_freq=args.target_network_update_freq,
train_freq=args.train_freq,
tot_steps=args.total_step_lim)
else:
'--freeze_layers',
type=int,
default=0,
help=
'Number of initial layers to freeze when fine-tuning | Choose from 1, 2, 3'
)
args = parser.parse_args()
logs_dir = 'logs'
if not os.path.exists(logs_dir):
os.makedirs(logs_dir)
args.save_dir = utils.get_save_dir(logs_dir, args.name)
# Training
if not torch.cuda.is_available() and args.cuda:
print(
'--cuda is passed but torch.cuda.is_available() returned False. Will use CPU instead.'
)
env = utils.wrap_deepmind(utils.make_atari(
args.env,
max_episode_steps=args.episode_length,
frameskip=args.frameskip),
frame_stack=True,
stacks=args.agent_history_length)
agent = Agent(env, args)
agent.train(args.episodes)
def _thunk():
env = make_atari('BreakoutNoFrameskip-v4')
env.seed(0 + rank)
return wrap_deepmind(env)