[BUG] Initializing any DL estimator with non-default values constructs these values only superficially

[achieveordie]

Describe the bug

A logical bug is present in all DL estimators and is caused by not passing custom parameters to its underlying Network since Network actually makes use of these values to create the DL model of appropriate size. Take a look at the example provided below.

To Reproduce

from pprint import pprint
from classification.deep_learning import CNNClassifier

model = CNNClassifier(
    kernel_size=3,
    n_conv_layers=1,
)

# model.summary() fails due to self.history being None, another bug

pprint(vars(model))
print("=="*25)
pprint(vars(model._network))

Output:

{'_X_metadata': [],
 '_class_dictionary': {},
 '_estimator_type': 'classifier',
 '_is_fitted': False,
 '_network': CNNNetwork(),
 '_tags_dynamic': {},
 '_threads_to_use': 1,
 'activation': 'sigmoid',
 'avg_pool_size': 3,
 'batch_size': 16,
 'callbacks': None,
 'classes_': [],
 'fit_time_': 0,
 'history': None,
 'kernel_size': 3,
 'loss': 'mean_squared_error',
 'metrics': None,
 'model_': None,
 'n_classes_': 0,
 'n_conv_layers': 1,
 'n_epochs': 2000,
 'optimizer': None,
 'random_state': None,
 'use_bias': True,
 'verbose': False}
==================================================
{'_tags_dynamic': {},
 'activation': 'sigmoid',
 'avg_pool_size': 3,
 'filter_sizes': [6, 12],
 'kernel_size': 7,
 'n_conv_layers': 2,
 'random_state': 0}

Expected behavior

This is happening because of self._network = CNNNetwork() (or corresponding Network) in the constructor of the estimator. The values that we receive are to be passed into the Network.

Additional context

I'm speculating this is the reason why the checks at #4036 are failing despite having very small hyperparameters. In any case, I'll open a PR to resolve this bug.

Versions

D:\Anaconda\envs\sktime-dev\lib\site-packages\_distutils_hack\__init__.py:33: UserWarning: Setuptools is replacing distutils.
  warnings.warn("Setuptools is replacing distutils.")

System:
    python: 3.8.13 | packaged by conda-forge | (default, Mar 25 2022, 05:59:00) [MSC v.1929 64 bit (AMD64)]
executable: D:\Anaconda\envs\sktime-dev\python.exe
   machine: Windows-10-10.0.22621-SP0

Python dependencies:
          pip: 22.3.1
   setuptools: 65.5.1
      sklearn: 1.1.3
       sktime: 0.15.0
  statsmodels: 0.13.5
        numpy: 1.22.4
        scipy: 1.9.3
       pandas: 1.5.1
   matplotlib: 3.6.2
       joblib: 1.2.0
        numba: 0.56.4
     pmdarima: 2.0.1
      tsfresh: 0.19.0

[fkiraly]

hm, this seems to be pretty systematic in the code base!

A baffling find, as this bug was hiding in plain sight. Kudos for spotting this.

What would be the best fix? Manually changing it all to passthrough of parameters?
Or, moving the neural network object to be a parameter of the estimator?

[achieveordie]

I'm leaning towards manually changing it for all estimators since it minimally changes the current structure/implementation.

Whenever we revise the architecture we can discuss the alternatives. What say?

[fkiraly]

Makes sense!