def simulate_daily_series(num_groups: int, num_timesteps: int, noise: float = 1.0) -> 'DataFrame':
# create realistic series:
tensor = _simulate(num_groups, num_timesteps, noise=noise, dt_unit='D')
# convert to dataset:
dataset = TimeSeriesDataset(
tensor,
group_names=range(num_groups),
start_times=[DEFAULT_START_DT] * num_groups,
measures=[['y']],
dt_unit='D'
)
# convert to dataframe:
df = dataset.to_dataframe()
# add predictors:
# TODO: meaningful predictors
df['X1'] = np.random.normal(size=len(df.index))
df['X2'] = np.random.normal(size=len(df.index))
# make number of timesteps per group non-uniform:
max_timestep_per_group = dict(zip(
range(num_groups),
np.random.choice(range(int(num_timesteps * .80), num_timesteps), size=num_groups, replace=True)
))
df['_max_time'] = DEFAULT_START_DT + df['group'].map(max_timestep_per_group)
df = df.loc[df['time'] <= df.pop('_max_time'), :].reset_index(drop=True)
return df
def simulate_daily_series(num_groups: int, num_timesteps: int, noise: float = 1.0) -> 'DataFrame':
season_spec = {
'season_start': np.datetime64('2007-01-01'), # arbitrary monday at midnight
'dt_unit': 'D'
}
# create realistic series:
tensor = _simulate(num_groups, num_timesteps, season_spec, noise=noise)
# convert to dataset:
dataset = TimeSeriesDataset(
tensor,
group_names=range(num_groups),
start_times=[season_spec['season_start']] * num_groups,
measures=[['y']],
dt_unit=season_spec['dt_unit']
)
# convert to dataframe:
df = dataset.to_dataframe()
# add predictors:
# TODO: meaningful predictors
df['X1'] = np.random.normal(size=len(df.index))
df['X2'] = np.random.normal(size=len(df.index))
# make number of timesteps per group non-uniform:
max_timestep_per_group = dict(zip(
range(num_groups),
np.random.choice(range(int(num_timesteps * .80), num_timesteps), size=num_groups, replace=True)
))
df['_max_time'] = season_spec['season_start'] + df['group'].map(max_timestep_per_group)
df = df.loc[df['time'] <= df.pop('_max_time'), :].reset_index(drop=True)
return df
def test_time_series_dataset(self):
values = torch.randn((3, 39, 2))
batch = TimeSeriesDataset(
values,
group_names=['one', 'two', 'three'],
start_times=[0, 0, 0],
measures=[['y1', 'y2']],
dt_unit=None
)
try:
import pandas as pd
except ImportError:
warn("Not testing TimeSeriesDataset.to_dataframe, pandas not installed.")
return
df1 = batch.to_dataframe()
df2 = pd.concat([
pd.DataFrame(values[i].numpy(), columns=batch.all_measures).assign(group=group, time=batch.times()[0])
for i, group in enumerate(batch.group_names)
])
self.assertTrue((df1 == df2).all().all())