[python-package] reorganize early stopping callback

[jameslamb]

Contributes to #3756.
Contributes to #3867.

Fixes the following errors from mypy:

callback.py:232: error: Too few arguments  [call-arg]
callback.py:308: error: Argument "train_name" to "_is_train_set" of "_EarlyStoppingCallback" has incompatible type "str | Callable[[Any, Any], list[Any]]"; expected "str"  [arg-type]
callback.py:398: error: Argument 3 to "_is_train_set" of "_EarlyStoppingCallback" has incompatible type "str | Callable[[Any, Any], list[Any]]"; expected "str"  [arg-type]

Those all come from the fact that when env.model in CallbackEnv is a CVBooster, any attributes not explicitly defined on the CVBooster class return a method that is called on each of the Booster objects in .boosters

LightGBM/python-package/lightgbm/engine.py

Lines 357 to 365 in 60a4a13

	def __getattr__(self, name: str) -> Callable[[Any, Any], List[Any]]:
	"""Redirect methods call of CVBooster."""
	def handler_function(*args: Any, **kwargs: Any) -> List[Any]:
	"""Call methods with each booster, and concatenate their results."""
	ret = []
	for booster in self.boosters:
	ret.append(getattr(booster, name)(*args, **kwargs))
	return ret
	return handler_function

For example:

import lightgbm as lgb
from sklearn.datasets import make_regression

X, y = make_regression(n_samples=10_000)
dtrain = lgb.Dataset(X, y)
results = lgb.cv(
    train_set=dtrain,
    params={"objective": "regression"},
    num_boost_round=7,
    nfold=3,
    stratified=False,
    return_cvbooster=True
)

cv_booster = results["cvbooster"]

cv_booster.num_trees()
# [7, 7, 7]

mypy is rightly complaining that given that, it isn't safe to treat env.model._train_data_name as if it was a string... since for a CVBooster, it won't be.

cv_booster._train_data_name
# <function CVBooster.__getattr__.<locals>.handler_function at 0x120c11bc0>

This PR proposes some changes to avoid those cases in the early stopping callback.

It also proposes some other changes to make the code in that callback a bit easier to understand (for example, calling _EarlyStoppingCallback._is_train_set() with keyword arguments).

[jmoralez]

Since this doesn't need the env we could move it to __init__, WDYT?

[jameslamb]

I agree! Having it be a loud error right when the callback is created, instead of deferred all the way til the first iteration of training, seems useful. And I'd be surprised to learn that there are other libraries or user code depending on initializing lgb.early_stopping() with a negative value of this and then somehow updating the value before the first time it's called.

Moved into __init__() in bd3366a.

[jameslamb]

Doing this broke this test:

LightGBM/tests/python_package_test/test_engine.py

Lines 4502 to 4507 in f175ceb

	def test_train_raises_informative_error_for_params_of_wrong_type():
	X, y = make_synthetic_regression()
	params = {"early_stopping_round": "too-many"}
	dtrain = lgb.Dataset(X, label=y)
	with pytest.raises(lgb.basic.LightGBMError, match="Parameter early_stopping_round should be of type int, got \"too-many\""):
	lgb.train(params, dtrain)

Now the error from the early stopping callback gets thrown before this one from the C++ side:

LightGBM/src/io/config.cpp

Line 37 in f175ceb

Log::Fatal("Parameter %s should be of type int, got \"%s\"", key.c_str(), candidate);

So I pushed 7a98d82, which:

• switches that test to use a different parameter, to keep covering that C++-side validation
• adds a test in test_callback.py on this specific error from lgb.early_stopping()
• adds an isinstance() check in the condition guarding that error in lgb.early_stopping(), so you can an informative error instead of something like TypeError: '<=' not supported between instances of 'str' and 'int'

Given all those changes, @jmoralez could you re-review? I don't want to sneak those in on your previous approval.

[jmoralez]

Love this

[jameslamb]

haha thank you, thank you 😂