river bridge: how do I use a Dense-Armor classifier in a river loop, and a river model after a Dense-Armor scaler?

river and Dense-Armor both learn one sample at a time, and both name the same methods: learn_one, predict_one, predict_proba_one. Because the interfaces match, the bridge does not convert anything: to_river(est) and from_river(est) return the object they receive. The two functions exist to say, in the code, which side the object is used from, and to raise a clear error when river is not installed.

The two steps below use a synthetic stream of 60 samples: 30 of class 0 with a near 0 to 3, and 30 of class 1 with a near 5 to 8.

1. A Dense-Armor classifier inside a river loop

to_river hands a Dense-Armor classifier to river code. The loop is the prequential one of river: predict first, then learn.

from river import metrics
from dense_armor.bridges.river import to_river
from dense_armor.utility.online.learn.online_classifiers import OnlineGaussianNB

s = [({"a": 0.1 * i, "b": 1.0}, 0) for i in range(30)]
s += [({"a": 5 + 0.1 * i, "b": 1.0}, 1) for i in range(30)]
m = to_river(OnlineGaussianNB())
acc = metrics.Accuracy()
for x, y in s:
    p = m.predict_one(x)
    if p is not None:
        acc.update(y, p)
    m.learn_one(x, y)
print(round(acc.get(), 4), m.predict_one({"a": 5.5, "b": 1.0}), m.predict_one({"a": 0.2, "b": 1.0}))

0.9661 1 0

The first sample has no prediction (the classifier has seen nothing), so the accuracy counts the other 59. After the stream, a point at a = 5.5 is class 1 and a point at a = 0.2 is class 0.

2. A river classifier after a Dense-Armor scaler

from_river hands a river model to Dense-Armor code. The | operator of Dense-Armor chains a scaler and the river classifier in one pipeline.

from river import linear_model
from dense_armor.bridges.river import from_river
from dense_armor.utility.signals.preprocessing.scale import StandardScaler

s = [({"a": 0.1 * i, "b": 1.0}, False) for i in range(30)]
s += [({"a": 5 + 0.1 * i, "b": 1.0}, True) for i in range(30)]
pipe = StandardScaler() | from_river(linear_model.LogisticRegression())
for x, y in s:
    pipe.learn_one(x, y)
print(pipe.predict_one({"a": 5.5, "b": 1.0}), pipe.predict_one({"a": 0.2, "b": 1.0}))

True False

The river logistic regression is binary and takes True and False as labels. The scaler of Dense-Armor standardises each feature before it reaches the river model.

Details

References

• river documentation, the learn_one and predict_one methods of its classifiers. Not read for this page: the two blocks above are the check that the interfaces match, on the installed version.

Choices stated explicitly

• The bridge is the identity. No wrapper class is written, because the objects already have the other side's methods. A wrapper would add a layer where the two interfaces are equal.
• The extra. pip install dense-armor[river] installs river>=0.26 on Python 3.11 or newer, as pyproject.toml declares. Importing dense_armor.bridges.river without river raises a ModuleNotFoundError that names the extra.

Measured cost

Not measured: the functions return their argument.

Not covered

River estimators that need a method Dense-Armor classifiers do not have, and Dense-Armor classifiers that need a method river does not call. The two blocks above run on river 0.26.1.