[python-package] Passing non-serializable loss function since 4.0.0.

[Deimos357]

Hello. I have custom non-serializable loss function which I was passing through fobj param of lgb.train().

Since #5052 it's not possible any more.

Also it's not possible to pass non-serializable loss function as params['fobj'] value, because of multiple copy.deepcopy(params) usages in train() function.

So what is the way to pass such loss function in 4.0.0?

[jameslamb]

Can you please share the function definition and a minimal, reproducible example showing how you're using it and the error you're getting?

[Deimos357]

@jameslamb
Minimal example for 4.0.0:

import threading

import lightgbm as lgb
import numpy as np


class LossWrapper:
    def __init__(self):
        self.barrier = threading.Barrier(2)

    def calc(self, pred, data):
        y = data.get_label()
        return pred - y, np.ones(len(y))


def main():
    dataset = lgb.Dataset(np.random.rand(100, 2), np.random.rand(100))
    loss_wrapper = LossWrapper()
    booster = lgb.train(params={'objective': loss_wrapper.calc}, train_set=dataset)


if __name__ == '__main__':
    main()

Output:

Traceback (most recent call last):
  File "D:\stuff\lgbm-test\400\main.py", line 23, in <module>
    main()
  File "D:\stuff\lgbm-test\400\main.py", line 19, in main
    booster = lgb.train(params={'objective': loss_wrapper.calc}, train_set=dataset)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\stuff\lgbm-test\400\venv\Lib\site-packages\lightgbm\engine.py", line 159, in train
    params = copy.deepcopy(params)
             ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 146, in deepcopy
    y = copier(x, memo)
        ^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 231, in _deepcopy_dict
    y[deepcopy(key, memo)] = deepcopy(value, memo)
                             ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 146, in deepcopy
    y = copier(x, memo)
        ^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 238, in _deepcopy_method
    return type(x)(x.__func__, deepcopy(x.__self__, memo))
                               ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 172, in deepcopy
    y = _reconstruct(x, memo, *rv)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 271, in _reconstruct
    state = deepcopy(state, memo)
            ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 146, in deepcopy
    y = copier(x, memo)
        ^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 231, in _deepcopy_dict
    y[deepcopy(key, memo)] = deepcopy(value, memo)
                             ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 172, in deepcopy
    y = _reconstruct(x, memo, *rv)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 271, in _reconstruct
    state = deepcopy(state, memo)
            ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 146, in deepcopy
    y = copier(x, memo)
        ^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 231, in _deepcopy_dict
    y[deepcopy(key, memo)] = deepcopy(value, memo)
                             ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 172, in deepcopy
    y = _reconstruct(x, memo, *rv)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 271, in _reconstruct
    state = deepcopy(state, memo)
            ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 146, in deepcopy
    y = copier(x, memo)
        ^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 231, in _deepcopy_dict
    y[deepcopy(key, memo)] = deepcopy(value, memo)
                             ^^^^^^^^^^^^^^^^^^^^^
  File "D:\python311\Lib\copy.py", line 161, in deepcopy
    rv = reductor(4)
         ^^^^^^^^^^^
TypeError: cannot pickle '_thread.lock' object

For 3.3.5 we can change
booster = lgb.train(params={'objective': loss_wrapper.calc}, train_set=dataset)
to
booster = lgb.train(params={}, train_set=dataset, fobj=loss_wrapper.calc)
and training will work fine.

Obviously this example makes no sense. In practice I'm synchronizing loss calculation of multiple boosters so they can share multi-variable loss through the wrapper.
I can rewrite whole python-package training loop to achieve the same results, but implementation with threading is much easier and cleaner.