[tabular][bugfix] fix tabular NN working with _ray_predict_oof

[liangfu]

Issue #, if available:

from autogluon.tabular import TabularDataset, TabularPredictor

# Training time:
train_data = TabularDataset('https://autogluon.s3.amazonaws.com/datasets/Inc/train.csv')
train_data = train_data.head(500)  # subsample for faster demo
print(train_data.head())
label = 'class'  # specifies which column do we want to predict
save_path = 'ag_models/'  # where to save trained models

if os.path.exists(save_path):
    shutil.rmtree(save_path, ignore_errors=True)

predictor = TabularPredictor(label=label, path=save_path). \
    fit(train_data, presets='good_quality', time_limit=120)

Training with presets='good_quality' will result in an error.

The fundamental reason is that _ray_predict_oof would delete fold_model.model, see

def _ray_predict_oof(fold_model, X_val_fold, y_val_fold, time_train_end_fold,
                     num_cpus=-1, save_bag_folds=True):
    pred_proba = fold_model.predict_proba(X_val_fold, num_cpus=num_cpus)
    time_pred_end_fold = time.time()
    fold_model.predict_time = time_pred_end_fold - time_train_end_fold
    fold_model.val_score = fold_model.score_with_y_pred_proba(y=y_val_fold,
                                                              y_pred_proba=pred_proba)
    fold_model.reduce_memory_size(remove_fit=True, remove_info=False, requires_save=True)
    if not save_bag_folds:
        fold_model.model = None
    return fold_model, pred_proba

Therefore, loading the PyTorch model with torch.load(path + "net.params") would result in an error.

Description of changes:

The proposed change is to include the PyTorch model inside the pickle file (model.pkl), just like other models.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Job PR-2650-23e1050 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2650/23e1050/index.html

[Innixma]

Interesting that this works now, I remember originally NN_TORCH wasn't serializable, but maybe that was NN_MXNET and we just kept that assumption when migrating. It could have also been some of the code that has since been removed that wasn't serializable. Either way, very nice that we can save it to pickle now.

[Innixma]

LGTM!