data = args.dataset
data = 'MT_001'
ts = np.load('./data/paper/{}.npy'.format(data))
ts = ts.reshape(-1)
# set_length = len(ts)
segmentation = int(len(ts) * 5 / 6)
# ts = ts.reshape(-1,1)
# scaler, ts_scaled = scale_raw(ts)
dim = args.dim
h = args.step
dim = 168
h = 24
args.model_name = '{}_h{}_model'.format(data, h)
dataset = create_dataset(ts, look_back=dim + h - 1)
scaler = MinMaxScaler(feature_range=(-1, 1))
dataset = scaler.fit_transform(dataset)
X, Y = dataset[:, :(0 - h)], dataset[:, (0 - h):]
train_data = dataset[:segmentation, :]
test_data = dataset[segmentation:, :]
# train_input = X[:segmentation, :]
# train_target = Y[:segmentation]
# test_input = X[segmentation:, :]
# test_target = Y[segmentation:]
# train_target = train_target.reshape(-1, h)
# test_target = test_target.reshape(-1, h)
params.merge(args)
params.dataset = 'london_2013_summary'
ts = np.load('./data/paper/{}.npy'.format(params.dataset))
ts = ts.reshape(-1)
# set_length = len(ts)
params.steps=168
params.H=24
params.cov_dim = 0
# test
params.num_epochs = 1000
params.model_name = '{}_h{}_convRNN'.format(params.dataset,params.H)
dataset = create_dataset(ts, look_back=params.steps + params.H - 1)
scaler = MinMaxScaler(feature_range=(-1, 1))
dataset = scaler.fit_transform(dataset)
# ts = unpadding(dataset)
# # ts_train = ts[:-9*24].copy()
# # ts_val = ts[-9*24-168:-7*24].copy()
# # ts_test= ts[-7*24-168:].copy()
kf = KFold(n_splits=params.k)
kf.get_n_splits(dataset)
params.restore = False
cvs = []
for i, (train_idx, test_idx) in tqdm(enumerate(kf.split(dataset))):
def pack_dataset(ts, steps, H):
_ = create_dataset(ts ,look_back= steps + H - 1)
X, Y = _[:, :(0 - H)], _[:, (0-H):].reshape(-1, H)
X = X[:,:, np.newaxis]
return X, Y
def pack_dataset(args):
json_path = os.path.join('models', 'deepAR.params.json')
params = Params(json_path)
params.merge(args)
ts = np.load('./data/paper/{}.npy'.format(params.dataset))
ts = ts.reshape(-1)
dataset_params_path = 'data/paper/lag_settings.json'
params.update(dataset_params_path)
params.steps = params.datasets[params.dataset]['lag_order']
params.train_window = params.steps + params.H
params.test_window = params.train_window
params.predict_start = params.steps
params.predict_steps = params.H
# params.num_class = 1
params.cov_dim = 0
# params.num_epochs = 30
dataset = create_dataset(ts, look_back=params.steps + params.H - 1)
tscv = TimeSeriesSplit(n_splits=params.k - 1)
*lst, last = tscv.split(dataset)
train_idx, test_idx = last
train_data, test_data = dataset[train_idx], dataset[test_idx]
train_tscv = TimeSeriesSplit(n_splits=params.k - 1)
*lst, last = train_tscv.split(train_data)
train_idx, valid_idx = last
train_data, valid_data = dataset[train_idx], dataset[valid_idx]
train_set, train_input = deepAR_dataset(train_data,
train=True,
h=params.H,
steps=params.steps,
sample_dense=args.sample_dense)
valid_set, valid_input = deepAR_dataset(valid_data,
train=True,
h=params.H,
steps=params.steps,
sample_dense=args.sample_dense)
test_set, test_input = deepAR_dataset(test_data,
train=False,
h=params.H,
steps=params.steps,
sample_dense=args.sample_dense)
test_target = test_input[:, params.predict_start:, 0]
train_input, train_v = deepAR_weight(train_input, params.steps)
train_sample = deepAR_WeightedSampler(train_v)
params.batch_size = 128
train_loader = DataLoader(train_set,
batch_size=params.batch_size,
sampler=train_sample,
num_workers=4)
valid_loader = DataLoader(valid_set,
batch_size=valid_set.samples,
sampler=RandomSampler(valid_set),
num_workers=4)
params.model_name = '{}_h{}_deepAR'.format(params.dataset, params.H)
params.model_dir = os.path.join('experiments', params.model_name)
params.plot_dir = os.path.join(params.model_dir, 'figures')
# create missing directories
os_makedirs(params.plot_dir)
i = -1
params.cv = i
logger = logging.getLogger('deepAR.cv{}'.format(i))
set_logger(os.path.join(params.model_dir, 'train.cv{}.log'.format(i)),
logger)
logger.info('Loading the datasets for batch-training')
params.restore = False
return params, logger, train_loader, valid_loader, test_input, test_target