A Fitted Object Should Remember Its Training

Fit Once, Apply the Same Preparation

BOTIE

Machine-learning libraries needed different algorithms to work together without every caller learning a new interface. The scikit-learn team’s 2013 API-design paper described a common arrangement: fit learns from a dataset, while prediction and transformation methods use the resulting state. That consistency lets a preprocessing object and a predictor participate in the same application.

For a centring procedure, the learned state includes the training mean. Later requests should use that stored reference. Relearning a mean from each request would give the same measurement a different meaning whenever its neighbours change. Separate fit and transform methods make the intended distinction visible, and a changed-batch test checks that the implementation actually respects it.

Sources for this history

Why Does One Visitor Change Another Visitor’s Result?

Botie centres sensor readings before predicting. A helper computes the mean of whatever batch it receives. That seemed convenient during training. In the service, the same pump now gets a different centred value depending on which other requests arrive beside it.

We need the fitted reference to belong to the trained object. Prediction reads that reference. It does not relearn it from the current visitors.

This stateless diagram models a saved reference; the Python object teaches its lifetime across calls.

Step 1 of 1

Two requests share one fitted reference

Training readings establish the centre. Later batches only use it.

The first prediction stays the same when the second batch has an extreme neighbour.
Stage	Readings	Centre used
Fit	[2,4,6]	4
Request A	[7]	4
Request B	[7,100]	4

Fit
Readings
[2,4,6]
Centre used
4
Request A
Readings
[7]
Centre used
4
Request B
Readings
[7,100]
Centre used
4

The first prediction stays the same when the second batch has an extreme neighbour.

Give the reference an owner

Only the fitting branch supplies the reference.

Keep the centre on the training branch

A new neighbour should not change a pump’s representation.

BOTIE

A new neighbour should not change a pump’s representation.

Choose an example

Fitting readings

1

2

3

Incoming request batch

1

Readings relative to the training mean

03

Follow the operations

1. Incoming request batch
1. 07
2. Fitting readings
1. 02
2. 14
3. 26
3. Mean
```
4
```
4. Subtract
1. 03
5. Result
1. 03

This stateless diagram models a saved reference; the Python object teaches its lifetime across calls.

1. Give the reference an owner

Make one Python object responsible for the mean learned by one transformer. An object created from a class is an instance, and a named value stored on it is an attribute. Here the mean is that attribute. Creating a second instance gives another model its own mean. A mutable global variable would let unrelated models overwrite each other.

Fit calculates the reference from training values. Transform checks that fitting happened and subtracts the stored reference from each incoming value.

Keep learned state separate from configuration. A setting says what fitting should do; an attribute records what fitting learned.

2. Follow a batch through the object

The training mean stays fixed while requests change. A NumPy array holds the numerical batch; the object owns the scalar reference.

training mean = (2 + 4 + 6) / 3 = 4
transform([7]) = [7 - 4] = [3]
transform([7, 100]) = [3, 96]

A request-local mean would centre the lone value to zero and change it again beside 100.

A request should not mutate its input

Subtraction creates a new array, leaving the stored readings available for another check.

Python

import numpy as np
raw = np.array([7., 100.])
centred = raw - 4
print(raw.tolist())
print(centred.tolist())

Output

[7.0, 100.0]
[3.0, 96.0]

The object in the challenge owns the reference. The request still owns its original array.

Check yourself

Which data may change the object?

Question 1 of 2

Which mean should transform use?

Use the information below.

Training [2,4,6]. Request [7,100].

3. Make invalid states hard to overlook

Before fitting, there is no learned centre. Raise a clear error instead of using an arbitrary zero. Reject empty or nonfinite training arrays. A deployment can load a previously fitted reference, but loading must check its schema and version.

Return a new transformed array. Changing the caller’s array in place makes later tests depend on call order.

Re-fitting is an explicit operation. Serving code should use an immutable snapshot or coordinate replacement so concurrent requests see a consistent version.

4. Test the object across calls

Compare a reading transformed alone with its result inside a batch. Check that the stored mean remains unchanged. Use a second instance to detect accidental shared state.

When a model is saved, save its fitted preprocessing too. Reconstructing a mean from today’s data creates a new model procedure even if the weights are unchanged.

These are software invariants. The best centre for the population remains an empirical modelling question.