User Guide Tutorial 03: Preprocessing › Scaling¶

This tutorial shows how to use TemporAI preprocessing.scaling plugins.

All `preprocessing.scaling` plugins¶

To see all the relevant plugins:

[ ]:

from tempor import plugin_loader

plugin_loader.list()["preprocessing"]["scaling"]

{'static': ['static_minmax_scaler', 'static_standard_scaler'],
 'temporal': ['ts_minmax_scaler', 'ts_standard_scaler']}

Now also load data source(s) we will use:

[ ]:

SineDataSource = plugin_loader.get_class("prediction.one_off.sine", plugin_type="datasource")

Using a static data scaling plugin¶

[ ]:

from tempor import plugin_loader

dataset = SineDataSource(static_scale=5.0, random_state=42).load()
print(dataset)

model = plugin_loader.get("preprocessing.scaling.static.static_minmax_scaler", static_imputer="mean")
print(model)

OneOffPredictionDataset(
    time_series=TimeSeriesSamples([100, *, 5]),
    static=StaticSamples([100, 4]),
    predictive=OneOffPredictionTaskData(targets=StaticSamples([100, 1]))
)
StaticMinMaxScaler(
    name='static_minmax_scaler',
    category='preprocessing.scaling.static',
    plugin_type='method',
    params={'feature_range': [0, 1], 'clip': False}
)

[ ]:

# Note the scale of static features.

from IPython.display import display

print("Min, max values per feature:")
display(dataset.static.dataframe().describe().T.loc[:, ["min", "max"]])  # type: ignore

dataset.static

Min, max values per feature:

	min	max
0	0.025308	4.818100
1	0.045985	4.950269
2	0.102922	4.952526
3	0.082939	4.858910

StaticSamples with data:

	0	1	2	3
sample_idx
0	1.872701	4.753572	3.659970	2.993292
1	0.780093	0.779973	0.290418	4.330881
2	3.005575	3.540363	0.102922	4.849549
3	4.162213	1.061696	0.909125	0.917023
4	1.521211	2.623782	2.159725	1.456146
...	...	...	...	...
95	0.590824	3.483686	3.144714	4.387360
96	3.675355	4.017405	1.410173	0.887198
97	3.753074	4.034174	4.952526	2.063088
98	1.860090	3.882065	1.704018	4.653787
99	4.292064	2.144970	3.754355	3.772714

100 rows × 4 columns

[ ]:

# Note the new scale of static features.

dataset = model.fit_transform(dataset)  # Or call fit() then transform().

print("Min, max values per feature:")
display(dataset.static.dataframe().describe().T.loc[:, ["min", "max"]])  # type: ignore

dataset.static

Min, max values per feature:

	min	max
0	0.0	1.0
1	0.0	1.0
2	0.0	1.0
3	0.0	1.0

StaticSamples with data:

	0	1	2	3
sample_idx
0	0.385452	0.959893	0.733472	0.609374
1	0.157483	0.149662	0.038662	0.889440
2	0.621823	0.712515	0.000000	0.998040
3	0.863151	0.207107	0.166241	0.174642
4	0.312115	0.525621	0.424118	0.287524
...	...	...	...	...
95	0.117993	0.700959	0.627225	0.901266
96	0.761570	0.809786	0.269558	0.168397
97	0.777786	0.813205	1.000000	0.414607
98	0.382821	0.782190	0.330150	0.957051
99	0.890244	0.427990	0.752934	0.772571

100 rows × 4 columns

Using a temporal data scaling plugin¶

[ ]:

from tempor import plugin_loader

dataset = SineDataSource(ts_scale=5.0, random_state=42).load()
print(dataset)

model = plugin_loader.get("preprocessing.scaling.temporal.ts_standard_scaler")
print(model)

OneOffPredictionDataset(
    time_series=TimeSeriesSamples([100, *, 5]),
    static=StaticSamples([100, 4]),
    predictive=OneOffPredictionTaskData(targets=StaticSamples([100, 1]))
)
TimeSeriesStandardScaler(
    name='ts_standard_scaler',
    category='preprocessing.scaling.temporal',
    plugin_type='method',
    params={'with_mean': True, 'with_std': True}
)

[ ]:

# Note the scale of time series features.

from IPython.display import display

print("Min, max values per feature:")
display(dataset.time_series.dataframe().describe().T.loc[:, ["min", "max"]])

dataset.time_series

Min, max values per feature:

	min	max
0	-4.999519	4.999999
1	-4.999982	4.999999
2	-4.999923	4.999992
3	-4.999979	5.000000
4	-4.999970	4.999928

TimeSeriesSamples with data:

		0	1	2	3	4
sample_idx	time_idx
0	0	-0.095075	-0.240884	-0.542729	2.209324	0.122539
	1	1.500152	1.822750	2.882952	4.450264	2.673609
	2	2.939520	3.567547	4.864963	4.930896	4.464728
	3	4.073484	4.688307	4.410788	3.461105	4.986786
	4	4.784229	4.988986	1.747861	0.622269	4.091392
...	...	...	...	...	...	...
99	5	4.835604	0.634449	4.634897	4.910111	4.815565
	6	3.532665	2.066645	2.845170	3.281070	4.946122
	7	1.263739	3.315606	0.121903	0.253339	4.998853
	8	-1.350749	4.270597	-2.641363	-2.882388	4.972930
	9	-3.595883	4.846944	-4.537960	-4.789380	4.868760

1000 rows × 5 columns

[ ]:

# Note the new scale of time series features.

dataset = model.fit_transform(dataset)  # Or call fit() then transform().

print("Min, max values per feature:")
display(dataset.time_series.dataframe().describe().T.loc[:, ["min", "max"]])

dataset.time_series

Min, max values per feature:

	min	max
0	-1.711349	1.200819
1	-1.724449	1.239101
2	-1.734762	1.230568
3	-1.592516	1.277314
4	-1.728804	1.177170