# create the training, val set, trend_set, status_dicts
train_set, val_set, extra_set = data.read_bundle_csv(
path="../data/14",
sep='\t',
filter_data=True,
fix_open_price=False,
percentage=0.8,
extra_indicator=True,
trend_names=['bollinger_bands', 'MACD', 'RSI'],
status_names=[])
env = environ.StocksEnv(train_set,
extra_set,
bars_count=BARS_COUNT,
reset_on_close=True,
random_ofs_on_reset=True,
volumes=False,
train_mode=True)
env = wrappers.TimeLimit(env, max_episode_steps=1000)
env_val = environ.StocksEnv(val_set,
extra_set,
bars_count=BARS_COUNT,
reset_on_close=True,
random_ofs_on_reset=True,
volumes=False,
train_mode=False)
# env_val = wrappers.TimeLimit(env_val, max_episode_steps=1000)
# create neural network
net = models.DoubleLSTM(price_input_size=env.price_size,
help="Name to use in output images")
parser.add_argument("--commission",
type=float,
default=0.1,
help="Commission size in percent, default=0.1")
parser.add_argument("--conv",
default=False,
action="store_true",
help="Use convolution model instead of FF")
args = parser.parse_args()
prices = data.load_relative(args.data)
env = environ.StocksEnv({"TEST": prices},
bars_count=args.bars,
reset_on_close=False,
commission=args.commission,
state_1d=args.conv,
random_ofs_on_reset=False,
reward_on_close=False,
volumes=False)
if args.conv:
net = models.DQNConv1D(env.observation_space.shape, env.action_space.n)
else:
net = models.SimpleFFDQN(env.observation_space.shape[0],
env.action_space.n)
net.load_state_dict(
torch.load(args.model, map_location=lambda storage, loc: storage))
obs = env.reset()
start_price = env._state._cur_close()
DEFAULT_VAL_STOCKS)
parser.add_argument("-r", "--run", required=True, help="Run name")
args = parser.parse_args()
device = torch.device("cuda" if args.cuda else "cpu")
saves_path = os.path.join("saves", args.run)
os.makedirs(saves_path, exist_ok=True)
if args.year is not None or os.path.isfile(args.data):
if args.year is not None:
stock_data = data.load_year_data(args.year)
else:
stock_data = {"YNDX": data.load_relative(args.data)}
env = environ.StocksEnv(stock_data,
bars_count=BARS_COUNT,
reset_on_close=True,
state_1d=False,
volumes=False)
env_tst = environ.StocksEnv(stock_data,
bars_count=BARS_COUNT,
reset_on_close=True,
state_1d=False)
elif os.path.isdir(args.data):
env = environ.StocksEnv.from_dir(args.data,
bars_count=BARS_COUNT,
reset_on_close=True,
state_1d=False)
env_tst = environ.StocksEnv.from_dir(args.data,
bars_count=BARS_COUNT,
reset_on_close=True,
state_1d=False)
args = parser.parse_args()
device = torch.device("cuda" if args.cuda else "cpu")
saves_path = SAVES_DIR / f"conv-{args.run}"
saves_path.mkdir(parents=True, exist_ok=True)
data_path = pathlib.Path(args.data)
val_path = pathlib.Path(args.val)
if args.year is not None or data_path.is_file():
if args.year is not None:
stock_data = data.load_year_data(args.year)
else:
stock_data = {"YNDX": data.load_relative(data_path)}
env = environ.StocksEnv(stock_data,
bars_count=BARS_COUNT,
state_1d=True)
env_tst = environ.StocksEnv(stock_data,
bars_count=BARS_COUNT,
state_1d=True)
elif data_path.is_dir():
env = environ.StocksEnv.from_dir(data_path,
bars_count=BARS_COUNT,
state_1d=True)
env_tst = environ.StocksEnv.from_dir(data_path,
bars_count=BARS_COUNT,
state_1d=True)
else:
raise RuntimeError("No data to train on")
env = gym.wrappers.TimeLimit(env, max_episode_steps=1000)