Model reuse is not possible when using the 'is_snap' option.

[yuuniee]

🐛 Bug

• Model reuse is not possible when using the 'is_snap' option.

To Reproduce

• I tried were:

  algo = ['denselight', ]
  automl_model = TabularAutoML(
  task = task, 
  timeout = TIMEOUT,
  cpu_limit = N_THREADS,
  gpu_ids='0',
  general_params = {"use_algos": [algo]}, # ['nn', 'mlp', 'dense', 'denselight', 'resnet', 'snn', 'node', 'autoint', 'fttransformer'] or custom torch model
  nn_params = {
      "n_epochs": EPOCHS, 
      "bs": TRAIN_BS, 
      "num_workers": 0, 
      "path_to_save": None, 
      "freeze_defaults": True, # False is Heavy...
      "cont_embedder": use_plr(USE_PLR), # 'linear' is better
      "cat_embedder": 'cat',
      "scheduler_params": { 'patience': 3, 'factor': 0.5, 'min_lr': 0.00001 },
      "opt_params": { 'lr': 1e-2,
                      'weight_decay': 0 },
      "drop_rate": 0.01,
      "is_snap": True,
      "snap_params": { 'k': 9, 'early_stopping': True, 'patience': 1, 'swa': True },
  },
  nn_pipeline_params = {
      "use_qnt": USE_QNT, 
      "use_te": False
  },
  reader_params = {
      'n_jobs': N_THREADS, 
      'cv': N_FOLDS, 
      'random_state': SEED, 
      'advanced_roles': ADVANCED_ROLES
  }, 
  )
  oof_pred = automl_model.fit_predict(df_train, roles = roles, verbose = 3, )

• Here, with the use of the {..., "is_snap": True,} option, reuse of the model becomes impossible and the resulting error code is as follows.

  self.best_loss = weights.pop("best_loss") KeyError: 'best_loss'

• You can check the log of the actual situation at the following link. (time line number: 92)

• https://www.kaggle.com/code/yuuniekiri/fork-of-home-credit-lightautoml-inference/log?scriptVersionId=183902165

• Model training was done in version 0.3.8.1, and model inference was done in 0.3.8.

Expected behavior

• No problem with inference even when reusing.

Additional context

• This is an error in the process of re-calling 'best_loss', which was removed with pop.
• I tried on 0.3.8.1 to retest and didn't find the issue. Maybe it only happens in 0.3.8.

Thank you.

Checklist

• [ ] bug description
• [ ] steps to reproduce
• [ ] expected behavior
• [ ] code sample / screenshots